Prepared for: Maienborn, Claudia, Klaus von Heusinger, and Paul Portner (eds).
Semantics: An International Handbook of Natural Language Meaning. Berlin:
Mouton de Gruyter.
Negative and positive polarity items:
Variation, licensing, and compositionality
Anastasia Giannakidou
University of Chicago
Draft, January 2008
In this chapter, we discuss the distribution and lexical properties of common varieties of
negative polarity items (NPIs) and positive polarity items (PPIs). We establish first that NPIs
can be licensed in negative, downward entailing, and nonveridical environments. Then we
examine if the scalarity approach (originating in Kadmon and Landman 1993) can handle the
attested NPI distribution and empirical variation. By positing a unitary lexical source for
NPIs—widening, plus EVEN— scalarity fails to capture the fact that a significant number of
NPIs are not scalar, and does not predict correctly NPI distribution in nonveridical contexts. It
also misses the variation within the scalar class between broader (any) and narrow NPIs
(either). Finally, scalarity predicts weaker effects (contradictions, presupposition failures)
with ill-formed NPIs than is actually the case. The variation approach (Giannakidou 1998,
2001, 2007), on the other hand, posits that besides scalarity, NPIs can be created because of
the presence of a variable in the NPI that cannot be interpreted deictically. This approach, by
allowing more lexical sources for NPIs, is consistent with the diversity of NPIs, and extends
easily to PPIs, which, as a class, appear to be non-scalar (Szabolcsi 2004 for some). Here one
can again advance the argument that PPI-status is due to various sources including
referentiality (some), and speaker commitment (with PPI adverbs; Ernst 2008).
1
Introduction: common paradigms of polarity items and basic terminology
Polarity phenomena in natural language are pervasive. In this chapter, we study the properties of
two of the most famous classes: ‘negative polarity items’ (NPIs) and ‘positive polarity items’
(PPIs). This first section offers the background necessary for our discussion, and introduces the
terminology that will help identify NPIs and PPIs as distinct among other PI-paradigms.
1. 1 Negative polarity items
NPIs are typologically very common (Haspelmath 1997 reports data from forty languages), and
seem to exist in virtually every language we consider. Their hallmark property is exclusion from
positive assertions with simple past (i.e. positive episodic sentences that make reference to a
single specific event; Giannakidou 1997). English any, ever (first identified in Klima’s 1964
work on English negation), Greek tipota, and Dutch ook maar iets are well-known NPIs:
(1)
a
b
Bill didn’t buy any books.
* Bill bought any books.
(versus: Bill bought {War and Peace/two books}).
1
(2)
a
b
* Bill has ever read War and Peace.
Bill hasn’t ever read War and Peace.
(3)
a
Dhen idhe
tipota o Janis.
not
saw anything the Johh
John didn’t see anything.
* Idhe tipota o Janis.
John saw anything.
b
(4)
a
b
Greek
Niemand heeft ook maar iets
nobody has
even
something
Nobody saw anything.
*Jan heeft ook maar iets gezien.
* John saw anything.
gezien.
seen
Dutch
Any, ever, ook maar iets, and tipota are ill-formed in positive episodic sentences, but become
fine with negation (or in habitual sentences, as we see later). NPIs belong to various syntactic
categories: there are nominal NPIs, but als NPI adverbs, NPI-verbs (hoeven, brauchen in Dutch
and German; van der Wouden 1994), NPI-EVEN particles (Greek oute, Giannakidou 2007;
Spanish nisiquiera Herburger 2003, a common category crosslinguistically), and English either:
(5)
(6)
(7)
(8)
a
b
a
b
a
b
a
b
(9)
a
b
Bill isn’t here yet.
*Bill is here yet.
I haven’t seen Bill in years.
* I saw Bill in years.
Bill doesn’t like pasta either.
* Bill likes pasta either.
Je hoeft niet te komen.
you need not to come
You need not come.
* Je hoeft te komen.
(Dutch)
Greek NPI-oute
Dhen theli na dhi
oute
to idhio tou to pedi.
not want3sg to see.3sg
even.NPI
the self his the child
He doesn’t want to see even his own child.
*Theli na dhi oute to idhio tou to pedi.
#He wants to see even his own child.
(Notice the contrast in (9) between Greek NPI-oute and English even in terms of status, a point to
which we return in section 6). NPIs are said to be licensed (or triggered; Ladusaw 1980) by
negation. Licensing normally says that the NPI must be in the scope of negation, and scope often
translates into overt c-command:
(10)
a
b
¬∃x. book (x) ∧ bought (b, x)
∃x. book (x) ∧ ¬ bought (b, x)
2
(11)
a
b
*Anydoby I didn’t see.
[Three students]F I didn’t see.
Not all NPIs are subject to overt c-command, and language specific conditions can play a role,
but it is certainly a strong tendency for (a subset of) NPIs to appear within the overt scope of
their licenser them. In this paper, our focus is on the semantic properties of NPIs, and details of
syntactic questions like overt c-command are put aside. In Giannakidou 1998, I offered a
semantic explanation for this tendency, by attributing it to the fact that NPIs are not referential:
construals like (11a) are not possible because any would have to be interpreted as a topic, but it
can’t (see Giannakidou 1998: 232-242 for details, including certain specificity effects). For an
attempt to reanalyze overt c-command as LF licensing, see Uribe-Etxebarria (1994).
1.2
Free choice items
Free choice items (FCIs) are often thought of in relation to NPIs. Because of any, which exhibits
both NPI and FCIs usage, attempts have been made to attribute NPIs and FCIs to a common
source (Chierchia 2006). We see in this paper that such attempts are bound to be unsuccessfull.
FCIs express what Vendler (1967) called freedom of choice, a property manifested in the
typical FCI use of any in generic sentences and with modal verbs:
(12)
(13)
Opjadhipote ghata kinigai pondikia.
(Greek)
Any
cat
hunts mice.
(English)
Opjosdhipote fititis
bori na lisi afto to provlima. (Greek)
Any
student
can solve this problem.
(English)
These are sentences about cats and students in general. Freedom of choice is described in terms
of domain widening (Kadmon and Landman 1993), indifference (von Fintel 2000, Horn 2000,
Giannakidou 2001), and is also known under the names of domain vagueness (Dayal 1995), nonindividuation (Tovena and Jayez 2005). These characterizations are intended to capture that freechoiceness comes with scalarity, contrasting with NPIs like tipota and ook maar iets which have
been described as non-scalar (Giannakidou 2008, Rullmann 1996 for Dutch).
Crosslinguistically free choice is marked with special morphology—e.g. Greek –dhipote,
in the FCI above (which is distinct lexically from NPI tipota). Many languages follow the pattern
of Greek with distinct lexicalizations (see Giannakidou and Cheng 2006, C. Lee 1997, and
references for more details). Importantly, FCIs are bad in both positive and negative sentences:
(14)
a.
b.
*Idha
opjondhipote.
saw.perf.1sg FC-person
‘*I saw anybody.’
*Dhen idha
opjondhipote.
not saw.perf.1sg FC-person
Intended: ‘I didn’t see anybody.’
(Greek; Giannakidou 2001)
Unlike NPIs, FCIs remain bad with negation in an episodic context. In Giannakidou 1998, 2001,
and Giannakidou and Cheng 2006, it is shown that anti-episodicity characterizes FCIs in many
languages, e.g. Spanish, Catalan, French, Chinese, and renders them polarity sensitive:
3
(15)
Spanish; Quer 1999
* (Non) Expulsaron del
partido a cualquier disidente.
not expel.3pl from-the party ACC FC
dissident
Intended: ‘*They expelled any dissident from the party.’
Intended: ‘They didn’t expel any dissident from the party.’
(16)
Catalan; Quer 1998
* (No) Li
va
comprar qualsevol ram.
not her/him aux.3sg to.buy FC
bouquet
Intended: ‘*S/he bought him/her any bouquet.’
Intended: ‘S/he did’t buy him/her any bouquet.’
But not all FCIs are polarity sensitive; e.g. German irgendein (Kratzer and Shimoyama 2002) is
fine in positive episodic contexts, and so is English wh-ever (Horn 2000, 2006):
(17)
a
(18)
b
a
b
Irgendjemand hat angerufen.
irgend-one
has called
* Opjosdhipote tilefonise.
Bill bought whichever book he liked.
*Bill bought whichever book.
(Kratzer and Shimoyama 2002: (6))
(Greek)
Indeed, looking across languages, one finds variation in FCIs (see Vlachou 2007 for French). I
am not going to discuss how FCIs come to be polarity sensitive here; rather, I use FCIs only to
compare to NPIs when it come to assessing the idea that they are d to the same source. Given
that (at least some) FCIs remain bad with negation, we must hypothesize that NPIs and FCIs, as
classes, cannot be sensitive to the same thing.
1.3
Positive polarity items (PPIs)
PIS, or POSITIVE PIS (PPIS) are expressions that are ‘repelled’ by negation and tend
to escape its scope. PPIs were first identified as a class in Baker (1970), and are discussed more
recently in Szabolcsi 2004, Nilsen 2003, and Ernst 2008. Expressions like some, already, would
rather, and speaker oriented adverbs have been identified as PPIs in the literature:
AFFIRMATIVE
(19)
(20)
(21)
(22)
a
b
a
b
a
b
a
b
Bill didn’t buy some books.
∃x. book (x) ∧ ¬ bought (b, x)
Bill would rather be in Montpellier.
# Bill wouldn’t rather be in Montpellier.
John is here already.
#John isn’t here already.
Unfortunately, John died.
# John didn’t unfortunately die.
As we see, some books exhibits the scope outside negation that any lacks, and only that. A
context where two books were bought by Bill and ten books were not, would verify John didn’t
4
buy some books but falsify John didn’t buy any books. PPIs like some are thus the opposite of
NPIs in terms of referential properties: they tend to be specific, and take wide scope, properties
to which we return in section 8.
Would rather and already are odd with negation, and likewise unfortunately. This
contrasting behavior of NPIs and PPIs prompted analyses of PPIs as anti-licensed by negation
(Giannakidou 1997, 1998, Progovac 1994; also Ladusaw 1979), but recently attempts have been
made to reformulate anti-licensing as a positive condition (Szabolcsi 2004, Giannakidou 2006).
1.4
Roadmap
Polarity phenomena are significant because they provide a case of well-formedness that is not
fully determined by syntax: the distribution of PIs is not fully predicted by their syntactic
category, and they seem to be sensitive to semantic factors of the sentence (roughly, the presence
or not of negation, or modality in the case of FCIs). This sensitivity addresses the fundamental
relation between syntactic and semantic well-formedness, and what impact, if any, pragmatics
can have on that relation. Polarity thus provides a uniquely complex domain to investigate the
relation between syntax-semantics and pragmatics, and the nature of semantic knowledge.
In this chapter, we examine the distribution and meaning of NPIs and PPIs, in the light of
the issues that have determined the research in the past 30 years: distributional variation (within
and across languages), mode of sanctioning (licensing, anti-licensing), and compositionality.
These issues are presented in section 2. In section 3, we examine the basic distribution patterns
of NPIs and see that NPIs are licensed in negative, downward entailing and nonveridical contexts
(which may be upward entailing). In section 4, we discuss the appearance of some NPIs in
veridical and non-DE contexts, and propose that this indicates that some NPIs are not actually
licensed, but rescued under certain pragmatic conditions. In section 6 we discuss the scalarity
approach and conclude that it is inadequate empirically and conceptually. In section 7, I discuss
the alternative I have been developing in my own work, where at least some polarity phenomena
are underlyingly referential dependencies. In my system, NPIs can also be created by variables
that cannot be used deictically or referentially. In section 8, we examine PPIs. We find here that
the scalar approaches have very little to say about PPIs and PPI-variation (Szabolcsi 2004, Ernst
2008). An approach based on veridicality, referentiality, and speaker commitment seems to be
more successful in this respect.
2
Core questions about polarity: sanctioning, variation, and sensitivity
NPI licensing has been a central issue in linguistic theory, and indeed one that has received
considerable attention since Klima's (1964) seminal work on English negation. In the earlier
works, the main goal has been English, but recent crosslinguistic studies have extended the
empirical domain of polarity, and made obvious a complexity that in the earlier works went
unnoticed. We now know that any is one of many PI paradigms in the world's languages, and
that the various NPIs are not subject to identical distributional restrictions.
The central theoretical task in most theories of polarity in the past 30 years has been to
delimit the set of potential NPI-licensers. This is known as the licensing question (Ladusaw
1996). In order to be able to predict if an expression can act as a licenser or not, we have come to
expect a coherent and relatively homogenous characterization of the set of expressions that can
5
possibly allow PIs within and across languages. There have been two main approaches to the
licensing question: the semantic, and the pragmatic negation-based one. The pragmatic approach,
best represented by Baker 1970 and Linebarger 1980, 1987, 1991 claims that all licensing is
done by negation, either by entailment (with negation), or (conventional or conversational)
implicature (when there is no negation). Linebarger also proposes that NPIs must be in the direct
scope of negation (Linebarger 1987), excluding PIs when “harmful” material intervenes between
negation and PI; these cases are known as intervention phenomena and will not be discussed in
the present paper for reasons of space (see Jackson 1994, however, for a good overview).
The semantic approach originates in Ladusaw 1980, and is modified in Zwarts 1995, von
Fintel 1999, and Giannakidou (1995, 1997, 1998, 1999, 2001, 2002, 2006). It maintains that the
set of possible NPI licensers must include expressions that are NEGATIVE, DOWNWARD ENTAILING
(DE, Ladusaw; Strawson DE; von Fintel 1999), and NONVERIDICAL, in this case often without
being negative or DE (Giannakidou, Zwarts). The extension to nonveridicality accounts for NPIs
in non-negative and non-DE contexts—questions, disjunctions, modalities, and directive
propositional attitudes such as want and hope. DE expressions are a subset of the nonveridical,
hence nonveridicality is an extension of DE that allows licensing by a non-DE expression.
A second important question concerns the status of illicit NPIs. Polarity, as mentioned
earlier, raises the question of well-fomedness that is not purely determined by syntax. If not
purely syntactic, then what is the nature of polarity ill-formedness? Since early on, the intuition
has been that sentences with failed NPIs, e.g. *Bill brought any presents, *Bill talked to John
either, are unacceptable in a sense stronger than mere lexical anomaly, or presupposition failure.
(23)
a
b
#The king of France arrived yesterday.
# Jason has a population of 3 million.
Here we have a presupposition failure and a category mistake (people cannot have populations),
and the sentences are perceived as odd, though we can almost figure out what they mean (e.g. If
Jason were a city, then it would have the population of 3 million, and likewise, If France had a
King then that person could have visited the museum yesterday). Ill-formed any sentences, on the
other hand, are odd in a qualitatively different way. But how exactly? 1
In the more than thirty years of research on polarity, with a few exceptions (Ladusaw
1996, Giannakidou 1998, 2001), scholars were eager to ignore the status question. Recently, the
tendency has been to view NPI ill-formedness as some kind of semantic/pragmatic murkiness,
with some NPIs being more murky than others along the same dimension (defined usually in
terms of some form of strength). This premise lies also at the foundation of approaches that seek
to reduce polarity to scalar deficiency (Lahiri 1998, Krifka 1995, Chierchia 2006); in these
approaches, NPI failures are typically reduced to presupposition failures or contradictions.
The desire to reduce all polarity failures, to a unitary source is, understandably, very
attractive to a semanticist. However, we saw already that FCIs and NPIs cannot plausibly be
reduced to the same source, as the latter improve, but the former do not, with episodic negation;
1
That polarity failures have stronger psychological status mere lexical anomalies is supported by recent processing
evidence (Saddy at al. 2004, Drenhaus et al 2006, Drenhaus, Blaszczak, and Schutte 2006). This research finds a
biphasic N400/ P600 pattern with violating conditions of NPIs, suggesting that the processing of NPIs is linked to
both syntactic and semantic factors: the N400s reflect the attempts to integrate the NPI semantically, and the P600
manifests the processing cost relating to syntax. These results are thus very challenging for approaches that try to
reduce NPI failures to mere presupposition failures or contradictions.
6
and there are NPIs that are not scalar (Greek tipota) yet exhibit distributions parallel to any,
which is. PPI indefinites are also not scalar (Giannakidou 1998, Szabolsci 2004). These and
more data (that we review in this paper) challenge the feasibility of the scalarity approach as a
general theory of polarity, and call for a variation account (Giannakidou 1998, 2001, 2007).
Status is linked to the question of compositionality: for a given PI-paradigm, what exactly
is responsible for the limited distribution we see? This question prompts a close examination of
the lexical contribution of individual classes of NPIs, but in the earlier days it was overlooked. In
those days, polarity conditions were postulated as global, composition external filters on
sentences (or grammatical representations), as can be seen in Ladusaw’s semantic filtering:
(24)
Semantic Filtering
(Ladusaw 1983)
grammatical (φ) =def Syn (φ) ∧ Sem (φ);
where Syn is syntactic well-formedness, Sem is semantic-well-formedness
Grammaticality is a conjunction of syntactic and semantic well-formedness. Syntactic rules
determine which structures will be grammatical and which not, and so do semantic rules. In the
case of polarity, a syntactically well-formed structure will be filtered out semantically: a sentence
with the NPI any is subject to a semantic licensing condition (that it requires negation), and this
condition is not satisfied if there is no negation. Why PIs are subject to such semantic conditions
remains a mystery: there is no obvious link between negation and the NPI in such an account.
In current theorizing, semantic filtering is unsatisfactory. We want to do away with a
polarity “module”, and understand how the lexical content of a given PI is responsible for its
limited distribution. In the new agenda, the polarity status is no longer stipulated but follows
compositionally from the lexical semantics, and PIs are no longer “special” or mysterious, but
manifestations of more familiar phenomena that exist in grammar anyway (scalarity, referential
dependencies). Recent examples of theories inspired by this goal are: Kadmon and Landman
1993, Lee and Horn 1994, Krifka 1995, Israel 1996, 1998, Lahiri 1998, Tovena 1998,
Giannakidou 1998, 2000, 2001, 2007, Chierchia 2006. A view of sensitivity has emerged where
PIs are ‘deficient’ because they contain additional or deficient components in their meanings,
thereby creating more demands on the environments of occurrence.
Having outlined these important foundational issues, we now go back to any and take a
closer look at its distribution.
3
Licensing and downward entailment
It soon became obvious that the distribution of any exceeds negation. I will proceed to show this
incrementally. First, any is fine in sentences with DE quantifiers (few professors), and in the
restriction of universal quantifiers such as every.
(25)
a
b
Every student who saw anything contacted the police.
{Few professors/*Many professors} invited any students.
The restriction of every, especially, makes a remarkable case against negation being the key
factor in the appearance of any. Any also appears in questions, and in if-clauses:
(26)
a
b
If you talk to any students, just let me know.
Did you see anybody?
7
The environments here are not negative, as noted already by Klima, who coined the term
“affective” to unify negation, questions, and the rest. Affective is just a “syntactico-semantic”
feature in Klima, but Ladusaw 1980 hypothesized that there is an underlying semantic property
characterizing all affective contexts: downward entailment (DE).
3.1
Downward entailment
Ladusaw proposed the following licensing condition for NPIs:
(27)
Ladusaw’s (1979) licensing condition
α is a trigger for negative polarity items in its scope iff α is downward entailing.
A trigger is an expression in the sentence whose presence is necessary in order to make a PI
legitimate; a trigger is also known as licenser. Unlike upward entailing (UE) functions, which are
order preserving and closed under supersets, DE functions are order reversing and closed under
subsets. Both are illustrated below (the definitions rely on Zwarts 1986, Kas 1993):
(28)
(29)
A function f is upward entailing iff for every X,Y: if X≤Y, then f(X) ≤ f(Y)
A function f is downward entailing iff for every X,Y: if X≤Y, then f(Y) ≤ f(Z)
UE functions support inference from sets to supersets, and DE functions support inference from
sets to subsets. In DE contexts, expressions denoting sets can be substituted for expressions
denoting subsets salva veritate. Negation, few students and the restriction of every are DE:
(30)
a
b
(31)
Lucy does not like linguistics.
[[syntax]] ⊆ [[linguistics ]]
____________________________
∴ Lucy does not like syntax.
Few students like linguistics.
[[syntax]] ⊆ [[linguistics ]]
____________________________
∴ Few students like syntax.
Every [student who likes linguistics] came to the party.
[[syntax]] ⊆ [[linguistics ]]
____________________________
∴ Every student who likes syntax came to the party.
Hence DE successfully captures the occurrence of NPIs with negation and in the restriction of
every, an environment that had nothing to do with negation. This was a remarkable result, and
DE initiated a fruitful research program for semanticists (Hoeksema 1986, Zwarts 1986, 1993,
van der Wouden 1994, Kas 1993, Dowty 1994, Lahiri 1998, among many others). One also finds
references to licensing environments as non-UE (Postal 2000, Progovac 1994). The shared
enthusiasm has been that we can characterize semantically the class of NPI-licensers. Yet this
advance brought with it a number of persistent limitations which we explore next.
3.2.
Problems with DE: NPIs appear in non-DE contexts
8
Critiques of DE are to be found in numerous places in the literature; I will mention here Zwarts
1995, Giannakidou (1997, 1998, 2006), and van Rooy (2003). The problems are empirical as
well as conceptual. One major problem with DE-conditions is, as mentioned earlier, that they are
not compositional: they are stipulated as global, composition external filters on representations
that contain PIs. Why NPIs are subject to licensing rules remains a mystery. In this section, we
further see that DE does not predict the correct set of licensing environments either in English or
crosslinguistically, and that it is not flexible enough to capture the variation attested.
3.2.1 Non-uniform licensing in the restrictions of universals
Though DE explains nicely the occurrence of NPIs in the restriction of every, it fails to capture
the contrast below between the restriction of every, which allows any, and the restriction of each
and both, which doesn’t (see Giannakidou 1997, 1998; also Horn 1972, Seuren 1984).
(32)
(33)
{Every student/ the students} who saw anything reported to the police.
a.
*/??Each student who saw anything reported to the police.
b.
*/??Both students who saw anything should report to the police.
This asymmetry is problematic for DE which predicts uniform licensing in universal restrictions.
The contrast follows if we assume that each and both, but not every, are presuppositional, and in
Giannakidou 1998, 1999 I suggested that presuppositional determiners are veridical:
(34)
A determiner δ is veridical iff δ NP V, uttered in c, presupposes that ∃x NP (x) in c;
otherwise, δ is nonveridical.
A veridical determiner such as each, both can only be defined if the domain is nonempty (or
contain exactly two members for both). This is what underlies the common characterization of
such determiners as D-linked. Every has no such restriction and can be used in a context that
allows an empty domain: We reviewed every linguistics candidate; that is, zero, since there
weren’t any is OK (for more details see Giannakidou 1998, 1999). The non-sanctioning with
each/ both suggests that, for NPIs, it is not DE that matters, but whether or not a nonempty
domain is presupposed.
3.2.2 NPIs in questions
DE cannot explain the occurrence of NPIs in questions. It is very hard to establish monotonicity
patterns in questions, and to my knowledge, there has been no successful attempt to do this, a
difficulty noted already by Ladusaw (see also Guerzoni and Sharvit 2007). This is a serious
empirical problem, as NPIs are very common in questions. In virtually every language that can
be identified to have an NPI with distribution that exceeds negation, the minimal extension is
questions, and not DE quantifiers. I illustrate below with Greek and Dutch:
(35)
a
Heb
je ook maar iets gezien?
have.2sg you anything
seen
‘Did you see anything?’
9
(Dutch)
b
Idhes
tipota?
saw.2sg anything
‘Did you see anything?’
(Greek)
The Dutch ook maar iets, and its Greek counterpart tipota, observed at the beginning to require
negation, are fine also in questions (and with no negative bias). Crucially, these NPIs are not
licensed by few or at most five (Zwarts 1981, Giannakidou 1998):
(36)
a
b
c
*Weinig mensen hebben ook maar iets gezien.
few
people have.3pl anything
seen
‘Few people saw anything.’
*/??Liji anthropi idhan
tipota.
few
people saw.3pl anything
‘Few people saw anything.’
* To poli 5 anthropi idhan tipota.
At most five people saw anything.
(Dutch)
(Greek)
The non-occurrence of ook maar iets and tipota with a DE quantifier, and their improvement
with questions (which are not DE), are problematic for the assumption that these items are
sensitive to DE. Notice also that even in English, native speakers may be hesitant about the
status of NPIs with DE quantifiers that are not readily ‘negative’—the judgment with at most,
especially when combined with higher number is quiet variable: ??/#At most 50 students
received anything has often been judged as odd by native speakers in my Semantics classes
(though of course more targeted research is need to establish this fact).
Finally, attempts to render ook maar iets and tipota strong NPIs (i.e. licensed by stronger,
negative licensers; Zwarts 1993, van der Wouden 1997) would fail too, because it cannot explain
their occurrence in questions. Ook maar iets, its temporal counterpart ooit (Hoeksema 1995), and
tipota are licensed by a wider array of environments that can be described as nonveridical.
3.2.3 Modal and other non-DE enviroments for NPIs
In addition to questions, there are many non-DE environments where NPIs are fine.
Observations about the individual cases can be found scattered in the literature, and here I begin
with NPIs in imperatives, with modal verbs, subjunctive/infinitival propositional attitudes,
habituals, and disjunctions (Giannakidou 1998, 1999, 2006).
(37)
a
b
c
Patise {kanena/opjodhipote} pliktro.
Press any key.
O Janis bori na milisi me {kanenan/opjondipote}.
John may talk to anybody.
O Janis ine prothimos na milisi me {kanenan/opjondipote}
John is willing to talk to anybody.
These environments are modal and non-monotone, and the occurrence of any here is unexpected
if DE is the licensing property. Notice that in Greek, both the non-scalar NPI and the scalar FCI
10
are good in modal contexts. Other cases that have been identified in the literature as problematic
for DE are the following. I illustrate below with any as well as the Greek NPIs and FCIs:
(38)
An kimithis me {opjondhipote/kanenan} tha se skotoso.
if sleep.2sg with FCI-person/NPI-person FUT you kill.1sg
‘If you sleep with anybody, I’ll kill you.’
Heim 1984 noted the problem with classic DE in conditionals.
(39)
Dhialekse {opjodhipote/kanena} vivlio.
(Imperatives)
choose.2sg FCI /
NPI
book
‘Choose any book.’
With kanena vivlio: ‘Choose some book or other.’
(40)
I Eleni dhiavaze sinithos
{kanena/opjodhipote} periodhiko.
the Ellen read.imperf.3sg usually NPI
magazine
Ellen used to read {some magazine or other/any magazine}.
(habitual)
Notice that in the habitual sentence Greek employs imperfective aspect. Often, the ambiguity of
the English past, which allows both perfective and imperfective readings, creates the illusion of
improvement of any, if speakers can render the sentence habitual in the context.
(41)
a
b
I bike mesa
kanenas
i
afisame to fos anameno.
either entered.3sg NPI
OR
left.1pl the light on
(??/#Either anybody came in OR we left the light on.)
*Bike mesa kanenas ke
afisame to fos anameno.
* Anybody came in AND we left the light on.
(disjunction)
(conjunction)
The occurrence of NPIs in disjunctions, quite unexpected under DE, was noted first in
Giannakidou 1998 (notice here an asymmetry with any, discussed in Giannakidou 1999, 2006).
The occurrence of kanena in disjunction follows if we assume that the NPI is licensed by
nonveridical operators: disjunctions are nonveridical, but conjunction is veridical (Zwarts 1995).
Recall, finally, that kanenas is non-scalar (Giannakidou 1998)—there is no even meaning
associated with it, nor does it make us think of a widened domain. Kanenas is simply a nonspecific existential, and I employed “some or other” here to indicate this. In the non-DE contexts,
then, scalar and non-scalar NPIs occur, and given this difference, it seems reasonable to expect
that scalar and non-scalar NPIs get sanctioned in these contexts for different reasons.
In sum, DE cannot provide the basis for unification of NPIs and NPI-licensers. We must
seek the unifying factor elsewhere; Zwarts 1995 and I suggested to explore nonveridicality.
11
4
Unifying NPI-licensers as a natural class: nonveridicality
In philosophy, the term veridicality is related to truth and sometimes existence; in Montague
1969, the verb see is called veridical because if I see a unicorn is true, then it must be true that a
unicorn exists; look for, on the other hand, is nonveridical because if I am looking for a unicorn
is true, it is not necessarily true that a unicorn exists. Giannakidou (1994 and sequel) and Zwarts
(1995) propose that polarity items are excluded from veridical sentences but are allowed in
nonveridical ones, and formalized definitions of veridicality based on truth. Truth and existence
are obviously related-- as can be seen more clearly in the discussion of nonveridicality of
determiners (Giannakidou 1998, 1998), subjunctive relative clauses, and the (non)-referentiality
of NPIs and PPIs I discuss in sections 7 and 8. 2
Veridicality is a property of sentence embedding functions: such a function F is veridical
if Fp entails or presupposes the truth of p. If inference to the truth of p under F is not possible, F
is nonveridical; nonveridicality thus captures a state of unknown (or as yet undefined) truth
value. Veridicality and nonveridicality thus replace the traditional characterizations of REALIS
(veridical) and IRREALIS (non-veridical)). Within the class of the nonveridical expressions,
negation is identified as ANTI-VERIDICAL in that NOT p entails that p is false.
(42)
(Non)veridicality for propositional operators
(Giannakidou 2006)
i. A propositional operator F is veridical iff Fp entails or presupposes that p is true in
some individual’s epistemic model ME(x); otherwise F is nonveridical.
ii. A nonveridical operator F is antiveridical iff Fp entails that not p in some individual’s
epistemic model: Fp → ¬ p in some ME(x).
(See also Bernardi 2002). Nonveridicality characterizes the meaning of functions that do not
ensure truth, e.g. negation, disjunction, volitional verbs want, suggest, insist, modal verbs and
adverbials, imperatives, questions, habituals, and the subjunctive. E.g. from the truth of John
wants to find a snake, or Please find a snake!, or Did you find a snake?, or John didn’t find a
snake, we cannot infer that John (or anybody else) actually found a snake. These are precisely
the contexts that we found above to license NPIs. Zwarts 1995 further presents a proof that DE
functions are a subset of the nonveridical, hence nonveridicality is not in competition bur rather a
conservative extension of DE, that affords a broader empirical coverage and strengthens the
semantic theory of NPIs in just the right way.
In the definition above, I use of a multimodel system where truth of a proposition is
evaluated with respect to an individual’s epistemic model, ME(x). ME(x) is a set of worlds
representing the epistemic status of the individual x, and relativization wrt ME(x) was necessary
(Giannakidou 1997, 1998, 1999) for explaining a contrast among propositional attitudes in NPI
licensing and mood selection that we examine next. The use of multiple models in assessing truth
is also emphasized in Tancredi 2007, where it is proposed that multiple models introduce also
multiple (but possibly overlapping) domains, a point to which we come back in section 7.
2
Lin 1996 suggests that non-existence is the relevant property for licensing NPIs in Chinese, a generalization which
Giannakidou and Cheng 2006 reformulate in terms of nonveridicality; for the related connection between irrealis
and nonveridicality, see Levinson 2006.
12
4.1
Propositional attitudes, NPIs, mood choice, and nonveridicality
Giannakidou 1995, 1998, 1999 observes a correlation between mood and NPI licensing in Greek:
NPIs appear in subjunctive complements (na) but not in indicative ones (oti), as shown below:
(43)
(44)
(45)
I Ariaδni epemine
na afiso {opjonδipote/kanenan} na perasi mesa.
the Ariadne insisted.3sg
subj let.1sg FC-person /NPI-person subj come.3sg in
‘Ariadne insisted that I allow anyone in.’
I Ariaδni θa iθele
na milisi me {opjonδipote/kanenan} fititi.
the Ariadne would like.3sg subj talk.1sg with FC- /NPIstudent
‘Ariadne would like to talk to any student.’
a
b
* O Pavlos pistevi
oti idhe me {kanenan/opjonδipote}.
the Paul believe.3sg that saw.3sg NPI / FCI
* Paul believes that he saw anybody.
* Kseri
oti
aγorasa
{kanena/opjoδipote} aftokinito.
know.3sg
that
bought.1sg NPI /
FCI
car
* He knows that I bought any car.
This correlation is not confined to Greek: any too is fine in the infinitival complements of would
like, insist, as we see, while it remains unacceptable in that complements of epistemic and factive
verbs. NPI-licensing thus correlates with mood choice: subjunctive and infinitival complements
allow NPIs, but indicatives do not (for data from more languages, including Romance and
Russian see Haspelmath 1997, Pereltsvaig 2000, Quer 1998).
Besides NPIs, what is also interesting here is the pattern of mood choice itself: why
would epistemic and verbs meaning dream/imagine select the indicative? This fact, commonly
observed in European languages (with the partial exception of Italian that allows both indicative
and subjucntive with epistemics; Portner 1992) is very surprising for the traditional assumption
that the indicative is the realis, and the subjunctive the irrealis mood (see also Farkas 1992). I
give below the two basic clusters of verbs in Greek:
(46)
Indicative verbs
assertives:
leo ‘say’, dhiavazo ‘read’, isxirizome ‘to claim’
fiction verbs: onirevome ‘to dream’, fandazome ‘imagine’
epistemics:
pistevo ‘believe’, nomizo ‘think’
factive verbs: xerome ‘be glad’, gnorizo ‘know’, metaniono ‘regret’
semifactives: anakalipto ‘discover’, thimame ‘remember’
(47)
Directive na- verbs ( equivalent to to-infinitivals)
volitionals: thelo ‘want’, elpizo ‘hope’, skopevo ‘plan’
directives:
dhiatazo ‘order’, simvulevo ‘advise’, protino ‘suggest’
modals:
(invariant) prepi ‘must’, bori ‘may’
permissives: epitrepo ‘allow’; apagorevo ‘forbid’ (negative permissive)
In Giannakidou 1995, 1998, 1999, I argued that all mood choice in Greek is regulated by
(non)veridicality (Borschev et al. 2007 extended this to Russian). Indicative selecting attitudes
13
are veridical. Consider the typical epistemic pistevo ‘believe’. Propositional attitudes express
relations between individuals (individual anchors in my system, following Farkas 1992), and
propositions. (In fact, every sentence is interpreted with respect to an individidual anchor, in the
default case the speaker). For α believes that p to be true, it must be the case that α, the main
clause subject, is committed to the truth of the embedded proposition p. Though the speaker
might disagree, a prerequisite for p to be true in (48) is that Jacob's epistemic model (i.e. the set
of worlds compatible with what Jacob believes) be a subset of the worlds where p is true:
ME(Jacob) ⊆ p. The speaker may believe or even know that what Jacob believes is false, but this
is irrelevant for Jacob’s beliefs.
(48)
[[ Jacob believes that Ariadne loves Paul]]c= 1 iff
∀w [w ∈ ME(Jacob) → w ∈ λw'. Ariadne loves Paul in w']
(See also Tancredi 2007 for a very similar formulation). Since all worlds in the model ME(Jacob)
are p-worlds, believe is veridical: [[pistevo (α, p ]] c = 1→ [[ p ]] MB(α)= 1; likewise, other epistemic
verbs such as think, and imagine, and fiction verbs (dream). Factives are strongly veridical: the
worlds in the speaker’s model too are p-words, consistent with the observation that factive
complements are presupposed to be true (see Giannakidou 1998, 1999 for more discussion).
The directive class, on the other hand, expresses a weaker relation between the speaker
and the embedded proposition: directive verbs do not require an individual’s commitment to the
truth of the embedded proposition. Consider thelo ‘want’. Intuitively, “wanting something is
preferring it to certain relevant alternatives, the relevant alternatives being those possibilities that
the agent believes will be realized if he does not get what he wants.” (Stalnaker 1984: 89). In
order to capture this, I use the subject's epistemic model again as the anchoring model (instead of
a buletic one, as in Portner 1997); and this model may be seen as including worlds representing
future realizations of the actual world (as suggested in Farkas 1992), designated as MEfut(su)
(though desires can also be about the past, but I ignore these cases here as they do not seem to
alter the overall picture). MEfut(su) interects with p, and is partitioned into two sets, W1 and W2.
W1 is the part that intersects with p. W2, is the remaining part and contains non-p worlds:
therefore W2 ∩ p = ∅. The worlds in W1 are more desired alternatives than the worlds in W2, but
from want (x, p) we cannot infer that p is true in MEfut(x). x wants p is true simply in case there is
a world in ME(x) that is a p-world:
(49)
[[ Jacob wants that Ariadne leave]]c= 1 if
∃w [w ∈ MEfut(jacob)∧ w ∈ λw'. Ariadne leave in w']
Hence directive verbs selecting typically to-infinitivals such as want, hope, and other volitional
and directives, are nonveridical. A similar analysis can be given for modal verbs (Giannakidou
1998). Nonveridicality thus helps us maintain the generalization that NPIs are admitted in
nonveridical contexts, and account for the core mood distinctions at the same time.
4.2
Strict NPIs: NPIs only licensed by antiveridicality
Within the class of of nonveridical functions, we identified antiveridal ones as those that entail
the negation of p. This category is relevant because it helps make sense of a very common
pattern crosslinguistically: NPIs that are licensed very narrowly and appear only with negation
14
and the antiveridical connective without (Giannakidou 1997, 1998, 1999). These NPIs are
known as “strong”, but I think it is more appropriate to refer to them as “strict”, as opposed to
any, kanenas which are ‘broad’ (rather than weak). Strict NPIs do not appear in nonveridical
environments that are not negative:
(a)
(50)
Either
a
b
c
d
(b)
(51)
Minimizers in Greek: Only allowed with negation (Giannakidou 1998, 1999):
Dhen dhino dhekara jia to ti th’apojinis.
not give.1sg damn
about the what will happen.2sg
I don’t give a damn about what will happen to you!
*Dhinis dhekara ja to ti tha apojino?
Do you give a dam about what will happen to me?
*An dhinis dhekara, tha me akousis.
(If you dive a damn, you’ll listen).
(52)
(53)
(c)
(54)
(Nathan 1999, Rullman 2003, Giannakidou 2006)
John didn't come either.
*Did John come either?
*I want John to come either.
* Pick this up either!
(OK: Pick this up too)
mo+ONE in Japanese (Nakanishi 2007, Yoshimura 2007)
a.
Watasi-wa
gakusei-o
{dare-mo / hito-ri-mo}
mi-nakat-ta.
I-TOP
student-ACC
{who-MO / one-CL-MO}
see-NEG-PAST
‘I didn’t see any students.’
b. *Gakusei-o
{dare-mo / hito-ri-mo}
mita-ra
siras-ero.
student-ACC {who-MO / one-CL-MO}
see-if
inform-IMP
‘If you see any student, inform me.’
Yoshimura 2007 argues that –mo lexicalizes as an item with ONE and gets special intonation, in
agreement with other strict NPIs crosslinguistically (e.g. Greek minimizers and n-words, as
argued in Giannakidou 1997, 1998, 2000); we review these in a minute.
(d)
(55)
NPI even (Giannakidou 2007; for Spanish, German see Herburger 2003, Schwarts 2005)
a
Dhen theli
na dhi
oute
to idhio tou to pedi.
not want3sg subj see.3sg even.NPI
the self his the child
He doesn’t want to see even his own child.
b
*Theli na dhi oute to idhio tou to pedi.
c
* Idhe oute to idio tou to pedi?
(e)
(56) a
b
(57)
a
N-words (Laka 1990) in strict negative concord languages (Giannakidou 2000, 2006)
* (Dhen) theli na dhi KANENAN.
Greek
Not want.3sg subj. see.1sg n-person
He doesn’t want to see anybody.
*(Non) ho visto nessuno.
Italian (Zanuttini 1991)
He didn’t see anybody.
*Idhe KANENAN?
15
b
Did he see anybody?
* Ho visto nessuno?
Did he see anybody?
N-words are typical cases of NPIs, since they need negation for legitimacy, especially in strict
negative concord languages. Space prevents me from expanding on this topic here, so I refer to
my earlier works for details. Importantly, n-words come often with special intonation, as
indicated above with upper case for Greek. I took the emphatic paradigm to be lexically distinct
from the non-emphatic paradigm we have seen so far. Emphatic intonation is systematically
licensed only under negation and xoris ‘without’, suggesting a clear split, observed also in
Japanese (Yoshimura 2007), as I mentioned earlier.
Besides negation, strict NPIs appear also in the scope of negative particles such as
without and its equivalents. Without p entails not p, hence without, just like negation, is
antiveridical (Giannakidou 1997, 1998):
(58)
a
b
c
…xoris na dhi oute to idhio tou to pedi.
without seeing even his own child.
…without talking to Bill either.
xoris na dhi KANENAN
Without seeing anybody.
We can then claim that the stricter class of NPIs is limited to antiveridical contexts only.
Crucially, the stricter NPI class is not licensed with weak DE quantifiers (at most n, few N):
(59)
a
b
*To poli pende fitites dhiavasan oute ena arthro.
‘At most 5 students read even one article.’
(Greek; Giannakidou 2007)
*To poli pende fitites dhiavasan KANENA arthro.
‘At most 5 students read any article.’
(For the relevant data in Japanese see Yoshimura 2007). The inability of non-negative DE
quantifiers to license strict NPIs prompted a distinction within the class of DE functions between
regular DE and more ‘negative’ functions: anti-additive and anti-morphic ones (Zwarts 1986,
van der Woudenn 1994):
(60) A function f is anti-additive iff f(a ∨ b) = fa ∧ fb.
(61) A function f is antimorphic iff it is antiadditive, and additionally f(a ∧ b) = fa ∨ fb.
An antimorphic function is equivalent to classical negation, and is antiveridical. An anti-additive
function is more negative that mere DE, but not fully negative (it reverses only disjunction).
Negative quantifiers (nobody, nothing) denote anti-additive functions—though it is also possible
that these can be analyzed as negations plus indefinites. The gradiance here is in terms of
‘strength’ of negation: strict (or strong’) NPIs require more negative licensers; antiveridicality
and antimorphicity are useful notions to define the sensitivity of this class.
Besides their severely restricted distribution in antiveridical contexts, strict NPI classes
are also distinct from the broader NPIs in that they obey syntactic locality restrictions: they are
16
only licensed by a clause-mate antiveridical expression. I emphasized this in Giannakidou 1998,
2000, and especially 2006, where examples are given with n-words from numerous languages
including Greek, Slavic, Hungarian, Romance (for the original data see also Progovac 1994,
Zannuttini 1991, Przepiorkopski and Kupc 1997, and others); in Giannakidou 2007 it is further
shown that NPI-oute is also licensed within a clause.
(62)
(63)
?? Dhen ipa oti o Janis diavase oute kan tis Sindaktikes Dhomes.
?? I didn’t say that John read even Syntactic Structures.
Dhen tu epetrepsan na diavasi oute kan tis Sindaktikes Dhomes.
They didn’t let him read even Syntactic Structures.
Notice the contrast with the broader NPIs any and kanenas, which can be licensed long distance
even through syntactic islands (Giannakidou and Quer 1995,1997, Giannakidou 1998, 2000):
(64)
a
b
Dhen tou ipan oti o Bill milise me kanenan.
They didn’t tell him that Bill talked to anybody.
Dhen prodose mistika pou eksethesan kanena.
He didn’t reveal secrets that exposed anybody.
(form Giannakidou and Quer 1997)
(relative clause)
The impossibility of long-distance licensing of strict NPIs, and the fact that it is observed
systematically in a number of (often unrelated) languages, suggests that with this class the need
to be in the semantic scope of their licenser is mediated by syntax. Locality can be implemented
as agreement: strict NPIs must have antiveridical features that must agree with the categorial
features of sentential negation dhen and without; agreement can only occur within a phase, hence
the locality (this is suggested in Giannakidou 2007). The stricter distribution to antiveridical
contexts follows, as there is no negative feature in nonveridical contexts that are not negative.
Alternatively, locality can be correlated with QR, as I argued in Giannakidou 1998, 2000 for
Greek n-words (an analysis extended further to Japanese by Shimoyama 2003, Yoshimura 2007;
and Hungarian, Suranyi 2006).
To sum up, we have seen in this section that the notion of nonveridicality allows us to unify the
polarity environments as a natural class, and predicts that NPIs may appear in contexts that are
unrelated to negation or DE as long as they are nonveridical: with modalities, directive
propositional attitudes, disjunctions, and questions. Antiveridicality, on the other hand, is the
notion we need as a criterion for the stricter NPI classes that are licensed narrowly by more
‘negative’ licensers. For this class, finally, we have seen that licensing is mediated by syntax,
e.g. as agreement (NPI-even) or QR (n-words and similar items).
17
5
NPIs in non DE and veridical contexts: two modes of sanctioning
In this section, we review unexpected occurrences of NPIs in veridical contexts, such as sentences
with only, emotive factive verbs, hardly, barely, and most. These cases were used by Linebarger
as an argument against the attempt to characterize semantically the class of NPI licensers. I start
with English minimizers, which as a class seem to behave more liberally than strict NPIs:
(65)
a
b
c
d
Ruth didn’t lift a finger to help me.
Ruth doesn’t give a damn about what I think.
Did Ruth lift a finger to help?
If you you give a damn, you’ll listen.
Minimizers are also fine with directive propositional attitudes, as is shown in the following data,
retrieved with Google, 10/17/2006; gratia Jason Merchant:
(66)
(67)
(68)
She’s still funny and cute and smart and I wish she gave a damn that we aren’t friends
anymore. I miss Candice. www.xanga.com/betweenIDs
“I just wish you gave a damn about something besides your television set.” Mr. Smith'
threw the remote control across the room stomped out of the room
...www.deadmule.com/content/word.of.mule.php?content_id=952
till the pianist finished, we left, and I dropped off tom and went home. Now I wish I had
said a word. It would have come out lame though, I just know it.
everything2.com/index.pl?node_id=1166781
Minimizers crosslinguistically do not form a natural class in terms of strict NPI status: English
ones are broad NPIs, but the Greek and Japanese minimizers are strict NPIs. English minimizers
and any are fine too in the scope of only, and in the complement clauses of emotive factive verbs.
(69)
a
b
c
d
I am glad he said a word!
I’m glad we got any tickets.
Mary regrets that she lifted a finger.
Only Mary {gives a damn/said anything}.
(from Kadmon and Landman)
These data are well known (see Atlas 1993 and Horn 1996), and pose a puzzle for both DE and
nonveridicality, since factives and only are veridical and not DE: 3
(70)
(71)
Only Bill left
→
Only Larry ate a vegetable
Bill left.
-/→ Only Larry ate broccoli.
3
Von Fintel (1999) and Hoeksema 1986 propose weaker versions of DE to deal with the
problem, by allowing the inference to the subset to be part of the common ground. For instance,
if we know in the context that John ate spinach, then from Only John ate a vegetable we can
infer that Only John ate spinach. By making this move, however, i.e. by allowing context
knowledge to influence reasoning, such weak DE overgenerates (Atlas 1993, Giannakidou
2006), and predicts that NPIs should occur also in positive sentences, thus failing to provide a
true explanation (see Giannakidou 2006 for discussion of more specific problems).
18
(72)
Larry may have eaten spinach.
Larry regrets that I bought a car.
-/→ Larry regrets that I bought a Honda.
Because, in fact, I bought a Ferrari, and Larry might not regret this at all.
The absence of DE in only and with emotive factives was used by Linebarger to claim that there
can be no semantic explanation for NPI-licensing at all. According to her, only and emotive
factives sanction NPIs because they convey the following negative inferences:
(73)
(74)
Nobody other than Larry ate spinach.
Larry wishes that I hadn’t bought a car.
For Linebarger, negation (as an abstract operator NOT) is the only true licenser of NPIs. In case
negation is not present in the sentence (i.e. the LF representation), she claims that NPIs can be
licensed derivatively by conveying a negative inference which is itself associated with an
appropriate LF representation: “the use of an NPI in a sentence whose LF does not license it
represents an allusion, one might say, to some entailed or implicated proposition, the negative
implicature, in which the NPI does occur in the immediate scope of negation” ( Linebarger 1991:
167). Such an allusion to negation further explains the following examples:
(75)
(76)
John hardly knows any syntax.
“Licensed” by: John does not know much syntax.
John barely knows anybody.
“Licensed” by: John doesn’t know many people.
Hardly and barely sanction NPIs in virtue of their conveying negative propositions. They are,
however, not DE (Horn 2002, Giannakidou 2002):
(77) a John barely studied linguistics -/→ John barely studied syntax
b John hardly talked to anybody -/→ John hardly talked to his mother
Likewise, the non-monotone exactly, and even non-DE most can sanction NPIs (the data are
taken from Israel 2004 who cites an earlier source):
(78)
Exactly two students said anything.
“Licensed” by: No more that two students said anything.
(79) a
Most children with any sense steal candy.
b.
Most people who would lift a finger to help Bill now are either very foolish or very
well-paid.
I am not going to criticize Linebarger’s theory at length here— see Giannakidou 1997,
and Israel 2004 for recent discussions. The main problem is that appeal to implicated negation
predicts unconstrained licensing, contrary to fact. Linebarger’s system predicts a direct
communication between syntax and pragmatics via LF, which hosts pragmatic information, an
assumption that we have reasons to believe is not correct (see my discussion next). The most
striking problem, however, is empirical: NPI sanctioning in these unexpected cases is not general
(Giannakidou 1997, 1998, 2006) but limited: it is allowed only with some NPIs.
19
Greek NPIs, either strict or broad, do not appear in these environments:
(80)
a
b
c
(81)
a
b
*Xerome pou dhinis dhekara.
I am glad you give a damn.
*/# Mono i Maria dhini dhekara.
Only Mary dives a damn.
#I Maria metaniose pou kounise to daktilaki tis.
Only literal interpretation: Mary regrets that she lifted her finger.
*Meta vias milise se kanenan.
He hardly talked to anybody.
* Akrivos dio fitites ipan tipota.
Exactly two students said anything.
The Greek NPIs need nonveridicality: they are ungrammatical with only, emotive factives, meta
vias ‘barely’ and exactly two (more data see Giannakidou 1997, 2006), as expected. Greek items,
then, are NPIs for which just implicatednegation does not suffice; such negation won't work with
strict NPIs in English either: *Only Bill saw John either, *Bill regrets that he saw Jane either
(see Giannakidou 2006 and Atlas 1993 for more illicit data with strong English NPIs).
In Giannakidou 2006 I built on the more marginal nature of this exceptional sanctioning,
and suggested that the variation can best be captured if we assume that any and minimizers
instantiate a class of NPIs that can either be licensed, or tolerated in a context if that context
gives rise globally to a nonveridical inference. In this case, the NPIs are rescued, not licensed:
(82)
Rescuing by nonveridicality (Giannakidou 2006)
A PI α can be rescued in the scope of a veridical expression β in a sentence S, if (a) the
global context C of S makes a proposition S' available which contains a nonveridical
expression β; and (b) α can be associated with β in S'.
“Association with a nonveridical proposition” means “be in the scope of a nonveridical
expression at a level other than LF”, however we are to define it, perhaps at the expressive
Emph-layer (suggested in Yoshimura 2007, building on Potts 2005). The global context C of S is
the set of propositions that arise from S without necessarily being entailed by it. C thus contains
the assertion (entailments), and presuppositions, implicatures. The negative proposition that is
responsible for rescuing will be conventionally contributed by some expression in the sentence.
In the case of only, it is the non-cancelable conjunct no x other than y P; with a negative factive,
e.g. regret, it is the counterfactual I wish that not p that is conventionally contributed by it. With
barely and hardly the NPIs are rescued via association with a background negative proposition
(whose precise status is still a matter of debate, see Horn 2002).
Rescuing builds on what I called indirect licensing in earlier work (Giannakidou 1998,
1999). Most importantly, rescuing happens in violation of scope at LF— an idea consistent with
Horn’s 2002 assertoric inertia, intended to capture precisely the state of affairs where an NPI
appears in the syntactic scope of an expression that does not have the semantic potential to
license it. The contrast between the Greek-type NPIs and English-type any/minimizers (see
Beaver and Clark 2003 for some data from Dutch) shows that we must allow NPI-rescuing as an
option in the grammar, but only as a secondary one: to my knowledge, there have been no NPIs
20
that appear to be rescued but not licensed (such NPIs would have to be fine with only and
emotive factives, for instance, but unacceptable with negation).
Given that the option of rescuing exists, languages may exploit it to a varying degree for
the various items. English seems to be more liberal than Greek in this respect. Ideally, one would
like to know why a particulat type of NPI favors rescuing, or why a given language X exploits
the rescuing strategy more liberally than a language Y, but I will leave this to future research.
This proposal has an important implication regarding LF and pragmatics. Contrary to
Linebarger, it becomes necessary to keep the LF "clean" of implicatures: if global information
were available at LF, it should be accessible to all PIs, thus making licensing possible, contrary
to fact. The empirical difference between licensing and rescuing can thus be taken as an
argument for the standard neo-Gricean view (pace Chierchia 2002, 2006; for a recent critique on
more general grounds see Russell 2007), and also in line with Potts (2005) where conventional
implicatures are computed at a level distinct from the truth conditional "at-issue" meaning.
6
Is scalarity enough for NPI licensing?
We now move on to address the question of compositionality: why do PIs appear in nonveridical
environments only? Why do they further partition the nonveridical space the way they do? Why
are some of them more narrow, and some more broad? Addressing this question is extremely
important, as I said at the beginning, because in answering it we gain a better understanding of
what NPIs mean, and how their meaning restricts their distribution. As a result, PIs will no
longer be mysterious but tractable.
There have been two approaches to compositionality. The first is the logician’s null
hypothesis: there is only one source of ill-formedness in polarity— scalarity, in some form or
other, is acknowledged as the culprit of polarity phenomena (Kadmon and Landman 1993,
Krifka 1995, Chierchia 2006, Lee and Horn 1994, Lahiri 1998). This position has been very
influential, so we need to assess very carefully what it says and how far it can go.
The second approach, which I have been advocating in Giannakidou (1998, 2001, 2006,
2007, to appear), is the variation position. This position acknowledges that (a) NPIs are not all
licensed in the same nonveridical environments, (b) not all NPIs are scalar or containing EVEN,
and (c) some NPIs are truly ungrammatical if not licensed (any, either, Greek kanenas, Dutch
ook maar indefinites, NPI EVENs), but others are not (some EVEN NPIs). This diversity makes
it more plausible to assume that the source of ill-formedness is not uniform, and we should
expect different accounts for different classes of PIs—though all of them consistent with why
NPIs require nonveridicality. In this approach, the obvious way to start is to say that PIs have a
lexical deficit (implemented via polarity features in Giannakidou 1997), and that this deficit
imposes restrictions on the contexts NPIs can be used. Nonveridical enviroments will have to be
harmless for the NPI-deficit. There are in principle three sources of lexical deficits: scalarity
(realized in lexically distinct EVEN-items; Giannakidou 2007), the presence of a dependent
variable (Giannakiodu 1998), and the presence of a syntactic feature that must be checked, thus
making the NPI subject to syntactic locality (as was shown to be the case of strict NPIs).
In this section, I address the scalarity approach and show why it will not be adequate. I
consider first domain widening, and then Lahiri’s implementation based on EVEN.
6.1
Domain widening: Kadmon and Landman and Chierchia
21
Kadmon and Landman’s 1993 quite influential paper on any proposed a unified theory for NPI
and FCI any by appealing to the notion of domain widening:
(83)
Meaning of any (Kadmon and Landman 1993)
any CN= the corresponding indefinite NP or CN with the additional semantic/pragmatic
characteristics (widening, strengthening) contributed by any.
(84)
Widening of any (Kadmon and Landman 1993)
In an NP of the form any CN, any widens the interpretation of the common noun phrase
along some contextual dimension.
The intuition behind domain widening is that any’s well-formedness depends on figuring out the
right domain—in this sense, any is just like any other quantifier. Any requires that its domain
must be extended (widened) along some contextual dimension. This sets any apart from other
quantifiers where typically the domain must be narrowed down.
Widening is coupled with strengthening:
(85)
Licensing condition for any: Strengthening
Any is licensed only if the widening that it induces creates a stronger statement, i.e. only
if the statement on the wide interpretation entails the statement on the narrow
interpretation.
Strengthening says that widening must have a purpose: it must make a stronger statement, a point
particularly central in Krifka 1995. Strengthening is satisfied in a negative context, but not in a
positive one:
(86)
a
b
I didn’t see any book on the table.
* I saw any book on the table.
The positive sentence is out because strengthening is not satisfied; in Chierchia’s words “domain
widening is pointless” (Chierchia 2006: 557) in a positive sentence. If it is true that I saw a book
in the widened domain, it is not necessarily true that I saw one in a (specific) narrower domain,
thus the widened statement is too weak, and for this reason not very informative. Here lies the
essence of all widening based accounts (see also Krifka 1995).
The key idea in Chierchia 2006 is the same—a unitary analysis for FCI and NPI any via
domain widening— only implemented in a system where implicatures project in a syntactic-like
manner. Chierchia excludes any in the positive sentence in the following way. First, the NPI
introduces alternative domains, indicated by the index i, which refers to numbers between 1 and
the maximum number (in this arbitrary case three) that we take our largest domain to consist of:
(87) a. *I saw any boy.
(Chierchia’s (47))
b. Meaning
∃w’∃x∈Dw’ [boy w’ (x) ∧ saw w (I, x)]
c. Alternatives
∃w’∃x∈Dw’ [boy i,w’ (x) ∧ saw w (I, x)], where 1 ≤ i ≤ 3
In a domain that consists of three boys, any boy makes us think of domains that contain 2 boys,
22
one boy, and all three boys. These alternatives are active and must be used to enrich plain
meaning, according to Chierchia. The domain of individuals is not ordered, but in choosing
among alternatives, speakers tend to go for the strongest one they have evidence for. In the case
above, we end up saying that even the most broad choice of D makes the sentence true: “in other
words, the base meaning will acquire an even-like flavor” (Chierchia 2006: 556).
The positive sentence also gives rise to the following implicature (Chierchia’s (48)):
(48) Implicature
∃w’∃x∈Dw’ [boy w’ (x) ∧ saw w (I, x)] ⊆c
∃w’∃x∈Dw’ [boy i,w’ (x) ∧ saw w (I, x)], where 1 ≤ i ≤ 3, and
p ⊆c q means: p is stronger (hence, less likely) than q relative to the common ground c
Chierchia claims that “given the way domains are chosen, (48) is logically false: all of the
alternatives are logically stronger than the statement in b; therefore, the latter statement cannot
be less likely than its alternatives. The positive sentence enriched by implicature (48) is
inconsistent, whence its deviance.” (Chierchia 2006: 556).
6.2
Problems with domain widening
Here I will take widening theories to task and see how far they can go in explaining (at least
some of) the core facts we saw in this paper, and how successful they are in being compositional.
In the widening approaches, the failure of any in a positive veridical sentence follows
from informativity, which is basic in conversation. Sentences with unlicensed any are claimed to
be impossible because the use of any renders them too weak to be informative (Kadmon and
Landman, Krifka), or inconsistent (Chierchia) after implicature enrichment. The first, rather
obvious, problem lies precisely here: the ill-formedness such a type of explanation predicts is
weak (Giannakidou 1998, 2001): sentences with failed any must have the same psychological
status as uninformative sentences. But this is not true. Speakers generally perceive a difference
between sentences that are ungrammatical, and sentences that are just odd; and uninformative
sentences—tautologies (The morning stat is the morning star), presupposition failures (The king
of France is my bother), contradictions (John is born in NY, and he is not born in NY), lexical
anomalies (The green ideas sleep furiously)— are never judged ungrammatical. The grammar
generates them, and speakers have a clear intuition that they can produce them; they can even
figure out ways to render them felicitous in certain contexts. Pragmatically odd sentences as thus
perhaps non-sensical in a default context, but certainly repairable and grammatically possible.
Positive sentences with any, however, (and the other kinds of NPIs we saw here), are
ungrammatical. The intuition has been that grammar does not generate these sentences, and
context manipulation does not have an effect on the NPI: *Any boy kissed Amy remains bad even
though we can still imagine a possible meaning of it (either some boy or every boy). This
difference is eagerly glossed over in the widening literature, but recent processing evidence
supports it (Saddy at al. 2004, Drenhaus et al 2006, Drenhaus, Blaszczak, and Schutte 2006, see
details in footnote 1), and challenges the fundamental idea of the widening approach that dooms
NPI failures to the realm of pragmatic uselessness. 4
4
Chierchia does acknowledge this weakness of the purely pragmatic account as the following
passage shows: “So why is a sentence like (47a) (an NPI-licensing violation) ungrammatical?
23
Hence widening alone cannot rule out NPIs in positive episodic sentences (Giannakidou
2001). A domain widened quantifier, e.g. in Every student that you can possibly imagine came to
the party is fine in a veridical context, though indeed some more thinking is required to figure
out what exactly the truth conditions of the sentence will be. But there is nothing fatally wrong,
or weak, in extending or even modalizing domains in positive veridical sentences. Any, unlike
every student that you can possibly imagine, is ill-formed not because it is domain widened, but
because it is subject to the composition external requirement of strengthening which works as a
filter. This recycles the semantic filtering of the earlier days, and means that the widening
approach is ultimately non-compositional.
A third problem lies with the very notion of widening: widening is not always present
with any, NPIs or FCIs (as shown also in Krifka 1995):
(88)
(89)
Pick any one of these 5 cards.
Consider any arbitrary number.
Here any extends over a very specific domain of the five cards in the context supplied by the
partitive. We cannot talk about domain widening in this case. Likewise, the set of numbers is
infinite, so it is hard to see what domain extension would yield in any arbitrary number.
A related, fourth, problem has to do with the fact that not all NPIs are scalar. Kanenas, as
mentioned earlier, is not:
(90)
Fere kanena gliko.
Bring some cake.
(Context: No need for something specific; it doesn’t matter what you bring really).
Kanena is a non-scalar existential and does not make us think of an extended domain; in fact, the
domain with kanena becomes somewhat irrelevant, as indicated above. Thus NPIs exist that are
non-scalar, yet unacceptable in the veridical positive sentence. For these cases at least, which are
in fact more common than we think (and labeled epistemically non-specific in Haspelmath 1997),
the domain widening analysis is not a plausible starting point.
Another problem arises with the fact that not all widened NPIs improve with negation:
any improves, but FCIs, in many languages, remain unacceptable. Recall:
(91)
a.
b.
*Idha
opjondhipote.
saw.perf.1sg FC-person
‘*I saw anybody.’
*Dhen idha
opjondhipote.
not saw.perf.1sg
FC-person
Intended: ‘I didn’t see anybody.’
(Greek; Giannakidou 2001)
There is an impasse here between the way domain widening explains the distribution of NPIs
(using Gricean principles) and the way such principles are typically taken to work….” (Chierchia
2006: 557). And later on, he posits a lexical entry for any (his (51)) where, in addition to
widening, any is claimed to have an uninterpretable feature [+σ] (Chierchia 2006: 559), ensuring
that any will be in the scope of some operator. It is checking of this feature that renders any
grammatical, a clear withdrawal from the purely pragmatic position.
24
Opjondhipote and all Greek FCIs are indeed scalar (Giannakidou 2001), as is typically the case
for FCIs crosslinguistically. The problem posed by these data is twofold. First, the nonimprovement of FCIs with negation undermines the idea that NPIs and FCI are due to the same
source (emphasized in Chierchia). Clearly, this cannot be the case, or at least it cannot generally
be the case. Second, the informativity based account predicts improvement with negation for all
widened items. The ill-formedness of opjondhipote then suggests that there is something other
than widening that further reduces the distribution of FCIs (Giannakidou 2001). Domain
widening falls short of deriving the correct set of facts.
The asymmetry within the class of NPIs and FCIs in terms of polarity poses a sixth
problem. Take any and whoever as a contrastive pair here. Both involve domain widening
(Jacobson 1995, Horn 2002, Giannakidou and Cheng 2006), but only any is polarity sensitive:
(92) a.
b
b.
Whoever saw a fly in his soup complained to the manager.
Irgendein hat angerufen.
*Anyone complained to the manager.
Likewise, irgendein in German patterns with whoever and not with any, despite its domain
widening (Kratzer and Shimoyama 2002). This variation leads to a conclusion similar to the one
we just reached regarding the variation with negation: there must be something other than
widening that further reduces the distribution of some scalar items, but not others.
The final puzzle for domain widening is posed by the fact that domain widened items like
any, as we saw in the previous subsections, are admitted in a large set of contexts not related to
negation, and some of them actually quite remote from it: modal verbs, imperatives, questions,
propositional attitudes, generic and habitual sentences. It is not obvious how the type of
explanation in terms of strength invoked by Kadmon and Landman, Krifka and Chierchia can be
extended to explain why NPIs are licit in these contexts. Consider, for example, the generic case:
(93)
Any cat hunts mice.
Here any cat, in the widened interpretation, does not entail the more narrow one, just like in the
positive veridical sentence—there are exceptions to generic generalizations (Krifka et al 1995).
Hence it follows that any should actually be unacceptable in this context. But it is not; rather,
genericity is a very common context for widened items.
Consider also NPIs in questions. I noted earlier that questions are not DE, and in the
informativity based theories that we are discussing this fact has been recognized (see especially
van Rooy 2003). The strategy within widening is to assume that NPIs are allowed in (nonrhetorical) questions either because they make the question more general than the corresponding
one without the NPI (Krifka), or because it turns a settled issue into an unsettled one (Kadmon &
Landman). But why should it a general question be preferred to a more specific one? And how,
or why, should this notion of generality be seen as a special case of strength in questions?
Van Rooy (2003) proposes that strength in questions must be reduced to entropy. Entropy
is the measure of the informativity value of a question. The informative value of question Q is
maximal just in case the answers to Q are all equally likely to be true. The value becomes less
than maximal when an NPI occurs (resulting in biased readings). In van Rooy’s words: “The NPI
weakens the satisfaction conditions for the positive answer, q, and strengthens the satisfaction
25
conditions for the negative answer ¬q.” (van Rooy 2003:xx). It is further proposed that entropy
can replace strength also in assertions, so it provides the unifying notion of strength.
However, this type of explanation is problematic. First of all, the general use of an NPI in
questions does not yield bias in polar questions—only the use of a certain type of NPIs does (and
is due to a particular kind of EVEN akin to at least; Giannakidou 2007). Second, the explanation
is analytically too weak: as I mentioned already, less informative sentences do not become less
grammatical. Van Rooy takes it for granted that informativity impacts grammaticality—
whether we want such a model may be a possible research question, and most likely to be
answered negatively; assuming a bona fide positive answer to it is not justified.
The pragmatic route in this direction produces, ultimately, liberal theories that predict more
fluid judgment than we have. Even any, an NPI of the more liberal kind as we noted, cannot
occur just as freely as van Rooy’s reasoning predicts. In fact, in questions, widened FC any,
along with FCI items crosslinguistically is systematically bad. The diagnostic is the ability to
modify any by almost (only FCI-any accepts this modification, Davison 1981):
(94) a
b
* Did you see almost anybody?
* Idhes sxedhon opjondhipote?
(Greek)
FCIs are prototypical domain wideners, but they fail to be licensed in questions (see also Quer
1998 for Spanish, Catalan data). This robust empirical fact is a surprise for van Rooy’s account.
Recall also that FCIs are generally bad with negation. This may support van Rooy’s attempt to
make negation and questions a natural class in terms of entropy/strength, but then the predicted
pattern is the opposite: items that involve domain widening are not admitted in negation and
questions. This suggests either that widening does not necessarily correlate with strength, or that
widening and strengthening (in whatever version) do not predict the right kind of polarity
sensitivity, a dooming conclusion in either case.
Notice also that if we unify negation and questions via strength, we can no longer
distinguish empirically between NPIs that are licensed by negation and questions (any) and those
that are licensed only by negation (either, Greek and Japanese minimizers, NPI even). By
identifying one culprit in both assertions and questions, we predict no contrast, but the contrast is
quite systematic across languages. It must be admitted, then, that we still need other factors
(syntactic or semantic) besides informativity for accurately restricting NPI distribution.
Finally, regarding other nonveridical NPI environments, it is unclear how informativity as
entropy would apply, e.g. in propositional attitudes, or with modal verbs, generic sentences. One
can of course stretch the notion, but then it runs the risk of becoming itself not very informative.
To summarize, I have shown in this section that domain widening, by appealing
exclusively to conversational principles such as informativity and strength, does not predict
either the correct distribution of various kinds of NPIs, or the correct effect of illicit NPIs. It also
misses the fact that not all NPIs are scalar, and is unable to account for variation with respect to
negation (negation does not save all widened items), or within widened items as regards their
polarity status or not (some FCIs are polarity sensitive, and some others aren’t). Finally, the
account was shown to be non-compositional, since strengthening is external to widening which
is the posited lexically property of any. In the end, it becomes obvious that even though widening
is indeed a lexical property of some NPIs (the scalar ones), when it comes to working out the
26
details of the explanation, it becomes impossible to make the argument that it is widening alone
that restricts the distribution of scalar NPIs.
6.3
Lahiri’s generalization
Lahiri 1998 formulates a widening account based on the idea that NPIs contain a low scalar
EVEN plus one. The empirical motivation came from Hindi NPIs (also used for FCIs), which
contain an EVEN—bhii—just like the Dutch ook maar, or the Japanese NPIs containing mo.
Lahiri 1998 argues that the low-likelihood presupposition of bhii creates a conflict when
combined with ONE in a positive sentence. The conflict, he argues, is resolved in negative and
DE contexts, thus NPIs will be admitted in these contexts.
The low-scalar property of English even is manifested in positive sentences, typically:
(95)
(96)
The Dean invited even Bill.
i. ∃x [x ≠ Bill ∧ invited (Dean, x)], and
ii. ∀x[x ≠Bill →likelihood (Dean inviting x) > likelihood (Dean inviting Bill)]
(95) asserts that the Dean invited Bill, and even contributes the presupposition (or conventional
implicature in Kartunnen and Peters 1977) in (96), i.e., that there is a set of alternatives to x, the
even phrase, that the context makes salient, and that these alternatives are ranked on a scale of
likelihood. The existence of alternatives is due to focus (Rooth 1985), observed with other focus
additive particles (too, also); low scalarity is a contribution of even (Horn 1989, Kay 1991). The
even phrase picks out the least likely individual(s) from the given set of alternatives. Likelihood
is understood as a possibility scale in the sense of Horn (1972, 1989). Low likelihood, when
combined with one in a positive sentence, yields oddity:
(97)
# Even ONE student arrived.
Notice that here we don’t talk about ungrammaticality. Even is odd because it is more likely that
one student came, not less likely. This is so because one is entailed by every other cardinality:
(98)
# ∃n [n ≠ one ∧ n students arrived] ∧ ∀n [n ≠ one →
likelihood (n students arriving) > likelihood (one student arriving)]
As a result, construals of EVEN one in positive sentences will always be problematic. With
negation, on the other hand, EVEN one will be OK, because, according to Lahiri, even scopes out
of negation and this yields the correct presupposition.
The appeal of Lahiri’s account is obvious: it remains faithful to the morphological facts
of Hindi and other languages that employ EVEN in some NPI formations. However, the link
between a low scalar EVEN and polarity is not as simple; nor does such link predict immediately
the appearance of NPIs in non negative nonveridical contexts, a fact that we also want to explain.
Lahiri predicts improvement of low scalar even with negation, assumes a straightforward
extension to DE, and overlooks all other contexts.
In Giannakidou (2007), I launched a detailed critique of the uniform low-scalar even
theories that need not be repeated here (see also Rullmann 1997, Hoeksema and Rullmann 2001,
for more critical points). The two main objections regarding Lahiri’s implementation are the
27
following. First, the improvement of even one and even NPIs with negation can have an
alternative analysis that does not raise even above negation. This analysis assigns to even a
presupposition of highest, not lowest likelihood, and was originally proposed by Rooth 1985.
(99) Bill didn’t invite even one student.
(100) ∃n [n ≠ one ∧ Bill didn’t invite n students] ∧ ∀n [n ≠ one →
likelihood (one student being invited) > likelihood (n students being invited)]
Even under negation can thus itself be an NPI, and it is very common crosslinguistically to find
designated lexical items EVEN with negation that carry a high-scalar presupposition (NPIEVENs, as I mentioned earlier). We must simply concede that English even and Hindi bhii, can
flexibly associate with low scalar and high scalar values, and that with negation only association
with high value is allowed, for reasons that we must explain (see Giannakidou 2007).
The second main objection, which I elaborate on next, is that empirically, it is not
desirable to assume either a continuum between negation and DE in terms of polarity even, or
that an EVEN meaning will necessarily improve with negation.
6.3. 1 With negation, low scalars do not always improve
In Giannakidou (2007) I discussed three items meaning even in Greek. In a positive sentence, as
shown below, only the low scalar akomi ke is used:
(101) a
I Maria efaje
akomi ke
to pagoto.
(positive EVEN)
the Maria ate
even
the ice cream.
b
*I Maria efaje
oute
to pagoto.
(NPI-EVEN)
the Maria ate
even
the ice cream
c
?#I Maria efaje
esto
to pagoto.
(flexible scale EVEN)
the Maria ate
even
the ice cream
(102) Presupposition of akomi ke
∃x [x ≠ ice-cream ∧ ate(Maria, x)], and
∀x [x ≠ ice-cream → likelihood (Maria eating x) > likelihood (Maria eating ice-cream)]
Akomi ke is just like even and Hindi bhii: it associates with the lowest end of a likelihood scale in
a positive sentence. And just like even and bhii, akomi ke is odd with one:
(103) #Akomi ke ENAS fititis irthe.
??Even ONE student arrived.
However, unlike even and bhii, akomi ke with one remains odd with negation, even if akomi ke
appears overtly above it (for more discussion see Giannakidou 2007):
(104) a
b
?#Akomi ke ENAS fititis
dhen irthe.
even one student
didn’t arrive.
?#Akomi ke enan fititi
dhen idha.
even one student I didn’t see.
28
The non-improvement of low scalar akomi ke with one is quite surprising for Lahiri who predicts
general improvement. But consider that one remains the highest likelihood also with negation:
(105) # ∃n [n ≠ one ∧ it is not the case that n students arrived] ∧ ∀n [n ≠ one →
likelihood (n students not arriving) > likelihood (one student not arriving)]
One is the most likely, not the least likely, cardinality and what we see here is a lexical
mismatch: a high scalar numeral like one cannot combine with an expression that must
associate with low scalar values like akomi ke. The contrast between Greek akomi ke, and
even/bhii then confirms that only the former is low scalar. For improvement of even/bhii
with negation something more needs to be said.
NPI-oute requires high likelihood; and one is fine with it:
(106) [[NOT oute (x) (P) ]] = 1 iff ¬ P (x) = 1;
∃y [y ≠ x ∧ C(y) ∧ ¬ P(y)] ∧
∀y [y ≠ x → likelihood (P(x)) > likelihood (P(y))]
(107) Oute ENAS fititis dhen irthe.
Not even one student arrived.
(assertion)
(presupposition)
Hence even/bhii can also associate with high values, and this is why they are fine with one.
Consider finally a third item, esto, also scalar, also odd with negation:
(108) ?# O Janis
dhen milise esto (ke) me tin Maria.
the John
not talked.3sg even with the Maria
John didn’t talk to even Maria.
(109) O Janis
dhen milise oute me tin Maria.
the John
not talked.3sg even with the Maria
John didn’t talk even to Maria.
Esto associates not with likelihood (though marginally it may be) but relies on the context to
provide a scale for it. I call it flexible scale even (see also Hoeksema and Rullmann for similar
facts in Dutch):
(110) [[esto (ke) (x) (P)]] = 1 iff P(x)= 1;
∃y [y ≠ x ∧ C(y) ∧ ¬P(y)] ∧
∃Qscalar [C(Q) ∧ ∀y [y ≠ x → Q(y) > Q (x)]]
(assertion)
(presupposition)
Esto is very close in meaning to at least—and it is the negative existential presupposition (absent
in at least, or at most an implicature) that renders it polarity sensitive. The non-improvement
with negation is due to the fact that the produced statement with esto is too weak: it merely says
that John didn’t talk to the least expected person— and this contradicts the choice to employ a
focus particle in the first place (Krirka 1995; for more details see Giannakidou 2007).
The main lesson here is that whether or not an NPI containing EVEN improves with
negation is a matter of lexical choice for that NPI, and is not generally predictable by lowlikelihood alone.
29
6.3.2 Even-NPIs are out with non-negative DE quantifiers
NPI oute and flexible scale esto do not improve with DE quantifiers (Giannakidou 2007):
(111) *To poli pende fitites dhiavasan oute ena arthro.
‘At most 5 students read even one article.’
(112) */??To poli pende pedhia efagan esto ena pagoto.
(?)At most five children ate
even one ice-cream
Greek
In fact, the positive low likelihood akomi ke is also odd:
(113) * To poli pende fitites aghorasan akomi ke ena vivlio.
? At most five students bought even one book.
Recall also the non-improvement with DE of the Dutch NPI ook maar ites, and Japanese daremo and hitori-mo that we saw earlier:
(114) *Weinigen zullen ook maar iets
few
will even
something
Few will achieve anything.
bereiken.
achieve
(Zwarts 1981).
Dutch
The non-improvement indicates that DE, at least in some languages, is not a sufficient condition
for the occurrence of EVEN ONE and EVEN. This challenges the continuum between negation,
DE and NPIs containing EVEN posited by Lahiri. NPI even itself is not licensed in contexts
other than negation and antiveridical without, as we saw ealrier, and this holds not just for Greek
oute, but NPI-even crosslinguistically. In Lahiri’s system there is no way to capture this fact (for
a similar point see Herburger and Mauck 2006 for data from Hebrew).
Finally, it is important to remember the difference in status between low likelihood
EVENs and NPI-EVEN. When illicit, the NPI oute is plainly ungrammatical; the other EVENs,
however, give rise to weaker effects— indicated by the use of #. Occasionally, these EVENs can
improve in a context if it is made consistent with their presuppositions somehow, a behavior
expected if what goes wrong is a presupposition clash, as I am arguing. The greater flexibility of
these items is due to the fact that their occurrence depends on a felicity condition (the satisfaction
of their presupposition). The rigidity of NPI-EVEN, on the other hand, is suggestive of a more
systematic “grammatical” sensitivity. It becomes necessary to distinguish between these two
different patterns, and I do not see a way to do this in the type of explanation proposed by Lahiri.
To sum up, we saw in section 6 that the accounts of NPIs based on sclalarity alone (widening,
EVEN-based) fail to predict the variable distribution, interpretation, and status of NPIs. They
posit a continuum between negation and DE which creates a handicap: it cannot capture the
contrast between strict NPIs with very narrow distribution in antiveridical contexts only, and
broader NPIs whose distribution is across nonveridical environments. We also saw that DE in
itself is actually a very marginal licenser. We must conclude then, that scalarity alone cannot
provide the basis for a general, and truly explanative theory of NPIs.
7
The variation approach to polarity: non-deictic variables
30
In Giannakidou 1998, 2001, 2007 I address the compositionality question from the perspective of
variation. In this perspective, polarity items are “created” because of distinctive lexical properties
they have. Some NPIs are scalar, but not all of them are. For scalar NPIs, I suggested a pragmatic
typology of four possible EVEN meanings, from which different analyses can be derived to
account for the variable patterns observed with EVEN-containing NPIs (Giannakidou 2007). I
will now turn to non-scalar NPI indefinites of the kanenas type. Such NPI indefinites exist in
many languages, and we saw that they are licensed broadly. Why are kanenas NPIs bad in
veridical sentences, and why do they improve in non-nonveridical sentences?
Let me start with what I said in Giannakidou (1998: 70-71, 139-40). I suggested there
that kanenas denotes a ‘dependent’ indefinite. A dependent existential is one that does not
introduce a discourse referent in the actual world (or in some individual’s epistemic model):
(115) An existential quantifier ∃xd is dependent iff the variable xd it contributes does not
introduce a discourse referent in the main context. (based on: Giannakidou 1998: 70)
A dependent existential is this peculiar “existential’ that cannot assert existence in a default
context. The intuition that some quantifiers cannot assert existence is also found in Matthewson’s
(1998) claim that some determiners (ku…a and kwelh…a in St’at’imcets Salish) “represent the
notion of ‘non-assertion of existence’” (Matthewson 1998: 179). (The Salish determiners appear
to be a bit broader in their distribution than the Greek kanenas—but it is important to note the
parallel). Matthewson further argues that such determiners do not entail non-existence of an
entity, rather they “merely fail to positively assert the existence of an entity” (Matthewson 1998:
179). In my system this intuition is formalized by using a designated variable: “xd” is a new kind
of variable—one that cannot introduce a discourse referent in a main context. Such variables
will not be able to be used in unembedded veridical sentences because they would be forced to
do exactly that. Under negation, there will be no problem:
(116) [[ kanenas ]] = person (xd)
(117) a
*Idha kanenan.
not saw.1sg anybody
b
Dhen idha kanenan.
not saw.1sg anybody
(118) a
b
[[ ∃xd person (xd) ∧ saw (I, xd) ]] g = undefined
[[ ¬ ∃xd [person (xd) ∧ saw (I, xd) ) ]] g = 1 iff no value a assigned to xd by g is
such that a is a person in c and I saw a.
Under negation, ∃-closure of xd will be fine because xd will not introduce a discourse referent.
Generally, then, dependent variables of this kind will be fine in the scope of nonveridical operators,
because the nonveridical operator is an embedding operator, and ensures that xd will not be forced
to introduce a discourse referent in the main context. den Dikken and Giannakidou 2002 extend
this idea to wh-the-hell phrases and treat them as dependent indefinites of this kind.
In Giannakidou 1998, I emphasized also that dependent reference does not imply lack of
reference:
(119) An dhis kanenan1,
pes tu1
na me perimeni.
31
If you see anybody,
tell him
to wait for me.
Here kanenan does introduce a discourse referent, which is subsequently modally subordinated
to the pronoun tu. Yet the introduction of the referent is done not in the default context but in an
embedded one (the protasis of the conditional). Dependent existentials thus receive values only
in embedded domains; their ‘deficiency’ is that the assignment function g cannot give them a
value in a main context. Two important assumptions here are that multiple domains are available
in a sentence when there is embedding (Giannakidou 1998, Tancredi 2007), just like there are
multiple models; recall our discussion of propositional attitudes. In fact, we can define
(following Tancredi) a conversion function that would assign a distinct domain to each model.
The second assumption is that the kanenas indefinite cannot receive a value in the main domain.
We can view this as inability to be interpreted in type e.
Dependent reference understood broadly says that some variables cannot be interpreted
referentially or deictically. These variables cannot be interpreted as free variables, and in order to be
well-formed, they must either be bound, refer back to antecedents, or be used in an embedded
domain. 5 Kanenas NPIs are referentially deficient in this way, and end up with limited
distribution in nonveridical contexts. In Giannakidou to appear, I analyze the Greek subjunctive as
containing a dependent temporal variable that cannot introduce a time in a main context;
temporal dependency is also limited to nonveridical contexts. In this program, polarity
phenomena are seen as a form of anaphoric like dependency; and if we look at the raw data, it
turns out that many NPIs are amenable to such an analysis.
The link between polarity and anaphora is not new. Progovac (1994) was the first to
exploit it, but her approach was purely syntactic: the goal was to transfer the binding principles
for pronouns to the locality conditions on NPIs. Another attempt with similar typological
concerns as mine is Schlenker 2003, 2004. In his system, there are two types of indexicals: the
unshiftable ones, i.e. those that can only be interpreted with respect to the context of the actual
speech act, and the shiftable ones, i.e. those that can be evaluated in this context as well as in the
context of the reported speech act. The driving premise in my program is exactly the same— that
not all variables are interpreted by the same procedures; then I argue that there is a natural class
of variables that are not deictic.
I will generalize, then, and propose the definition below for dependent variables:
(120)
Dependent variables
An variable xd is dependent iff xd cannot be interpreted as a free variable.
From this it follows that a dependent variable cannot be linked to a discourse referent. In order to
become licit, such variable has three options: (a) binding by a higher quantifier; (b) embedding
under negation and noveridical operators; (c) co-reference (a strategy exploited largely by the
subjunctive in embedded contexts).
In Giannakidou 2001, I argued that FCIs contain a dependent world variable wd. This
variable can only become licit via binding. In an episodic context (positive and negative) there is
no operator to bind it, the variable remains unbound, thus ill-formed:
5
Farkas 2002 uses the term ‘dependent’ in a narrower sense: to refer to variables that are well-formed only if be
bound by higher quantifiers (e.g. Hungarian indefinites egy-egy which only occur as bound existentials). I envision
the class of dependent variables broadly, as a label for the whole range of narrow scope variables— including the
ones studied by Farkas.
32
(121) [[ opjosdhipote]] = person(x)(wd)
(122) *Idha opjondipote
not saw.1sg anybody
(123) *Dhen idha opjondipote
not saw.1sg anybody
(124) [[ ∃x [person (x, wd) ∧ saw (I, x)] ]] = undefined in c because wd can’t get a value in c
(125) [[ ¬∃x [person (x, wd) ∧ saw (I, x)] ]] = undefined in c because wd can’t get a value c
The variable wd is non-deictic, and can never receive a default value to refer to the actual world.
This derives the non-improvement of FCIs with episodic negation—without a binder, wd remains
illicit in both positive and negative sentences—and predicts the correct distribution of FCIs in
non-episodic contexts where will be abinder for wd. (The universal effect is further derived by a
presupposition, not pertinent to our discussion here). An illicit (unbound) variable, in this
reasoning leads to ungrammaticality, just like *The Pope saw herself is ungrammatical.
The notion of dependent reference, as conceived of here, allows us to unify the class of
indefinites that are always interpreted with narrow scope. Dependent reference, naturally, is not
limited to polarity, but extends to other expressions that we didn’t think of NPIs before— bare
plurals in their existential use, bare singulars in English, Romance and Greek, and incorporated
object nouns in languages that allow them. These items are unusable if forced to refer to a particular
object in the highest domain. Hence, the notion of dependent reference embeds polarity into the
larger class of non-specificity phenomena, a quite welcome result, as it explains the polarity status
as a manifestation of something more familiar.
Before closing, let me raise some questions for further research within this program. The
first is how to refine and formalize further the property of dependence. I have done this here
following Giannakidou 1998 by using designated variables (subscripted by d), and treated the
dependency as a condition on the variable assignment function (Giannakidou 2001, Giannakidou
and Cheng 2006): that g cannot assign a value to xd in an unembedded context. The difference, in
particular, between NPIs and FCIs vis a vis negation suggests that we need to further distinguish
between dependent variables that can only become licit through binding (FCIs), and those who
just cannot refer to individuals in the domain DM of the main context (kanenas). The former will
not improve with negation because negation is not a binder, but the latter will, because they will
be existentially closed under negation, and will not refer to an individual in DM.
The inability to receive a value at DM deserves some more thought. If DM contains the
actual individuals in the context, does this mean that non-deictic indefinites with this deficiency
cannot refer to actual individuals at all? I don’t think so. As I mentioned briefly earlier, the
various domains can be overlapping, and clearly, under negation, xd can indeed receive values
from DM. This shows that what is troublesome for xd is that xd cannot be linked by default to a
value in DM, and this is what I tried to capture by calling this variable non-deictic. More work is
need in order to establish how exactly sortal differences of the kind I am talking about are
encoded in the grammar, but the main message is that we need such differences.
Another question is: how does my notion of non-deictic variable relate to Partee’s 1986
theory of flexible NP-interpretation? Given that dependent indefinites are always non-specific
and narrow scope, and given that such indefinites are recently thought as denoting properties (et)
(McNally and others), can the difference between “regular’ indefinites, which are freer in the
referential and scope properties, and dependent ones, be reduced to a type difference? Is it plausible
33
to say that dependent indefinites are property-type only? I am not sure that a positive answer can be
given to this question, since dependent indefinites can indeed refer to individuals in embedded
contexts (Giannakidou 1998, 2001). However, it is worth exploring whether the sortal difference I
propose here can be reduced to a more familiar type difference.
Finally, if what underlies polarity indefinites is an inability to refer to individuals in a
default context, the relevance of notions such as non-existence (Lin 1996) is not accidental.
(Non)existence and (non)veridicality are tightly connected: in a veridical context you are forced to
refer, thus exist; in a nonveridical one, you are not.
8
Positive polarity items
Positive polarity items (PPIs) are thought to have “the boring property that they cannot scope
below negation” (Szabolcsi 2004: 409). In this section, we consider two representative members
of the class—the indefinite some, and speaker oriented adverbs (Nilsen 2003, Ernst 2008). Our
starting point, as with NPIs, will be that PPIs do not form a homogeneous class. We follow
Szabolcsi’s conclusion that “the someone/something-type PPIs share properties with NPIs”
(Szaboclsi 2004: 409), but we show this to hold not because some with negation is an NPI, as
argued by Szabolcsi, but because some indefinites have nothing special, and they can generally
scope below any operator, regardless of negation. What is special, is the fact that someindefinites are accented under clausemate negation, and it is only in this use that they are PPIs.
The accent signals scoping above negation, a pattern that we find in other languages (Greek,
Giannakidou 1998, 2000; Japanese, Yoshimura 2007) when quantifiers scope above negation.
Following Ernst, on the other hand, the distribution of speaker oriented adverbs will be
argued to be sensitive to speaker commitment: strong evaluative adverbs like unfortunately are
factive, and express full speaker commitment to the truth of p. In Ernst’s terms, this makes them
subjective, and therefore felicitous only in veridical contexts, which guarantee truth of p in all
worlds in the speaker’s belief model (as we saw). This explains the inability of speaker oriented
adverbs to occur in nonveridical contexts such as questions, conditionals, etc. As we move to less
subjective adjectives, e.g. possibly and obviously, the truth condition is weakened, (p need not be
true in all worlds in the speaker’s model), and this results in compatibility in some nonveridical
contexts. Both Szabolcsi and Ernst show that scalarity plays no role in PPI distribution.
8. 1 Two types of some indefinites: emphatic and non-emphatic some
Ever since Jespersen, some has been thought of as an PPI in that it must scope above negation:
(126)
You didn’t see something.
This sentence cannot mean that you didn’t see anything, where an existential quantifier scopes
below negation. Scoping above negation is the defining property of PPI-hood, and it is indeed
observed with equivalent items across many languages, e.g. Serbocroatian (Progovac 1994,
2000), Dutch (van der Wouden 1994), Greek (Giannakidou 1997, 1998), Hungarian (Toth 1996,
Szabolsci 2004), among others. It has gone unnoticed, however, that this scoping has a particular
intonation: some is accented (uppercase henceforth), and negation is de-accented:
(127)
You didn’t see SOMETHING.
34
The reverse pattern, when some is de-accented, allows, and perhaps even prefers, a narrow scope
reading under negation. In this case, negation is accented (emphatic denial). PPIs like some are
thought to be “allergic” to negation, and this allergy translated as being anti-licensed by negation
(Progovac, Giannakidou, Ladusaw) in the sense that some must raise structurally in a position
above negation (the exact nature of which is immaterial here). PPIs in this context are the reverse
of NPIs, and scholars thought of them as contrasting pairs. 6
The need to escape the scope of negation has been lexically motivated by assuming that
some indefinites are specific (Progovac) referential—they always assert existence in some model,
as I suggested in Giannnakidou (1997, 1998). In either case, they get existentially closed above
negation only. Apart from clausemate negation, some is excluded from the immediate scope of a
negative quantifier and other anti veridical operators (without):
(128)
a
b
c
John didn’ t call SOMEONE.
# not > some
Nobody called SOMEONE.
# no one > some
John came to the party without SOMEONE. # without > some
(The non-emphatic versions above are fine with narrow scope some, a point to which we return).
So we can generalize that emphatic SOME must scope above antiveridical elements. However,
some-PPIs have been notorious for allegedly scoping below non-local negation:
(129)
Bill didn’t say that you saw something.
not > say> some
This narrow scoping wrt negation is peculiar for items that must escape the scope of negation in
the first place. To make things worse, narrow scoping of some is observed even with local
negation, if negation + some is found under an NPI-trigger—a fact noted in Jespersen, Baker and
Postal, and emphasized by Szabolsci. I give below data from Szabolcsi (2004: (33)-(40)):
(33)
(34)
(35)
(36)
(37)
(38)
(39)
(40)
I don’t think that John didn’t call someone.
No one thinks that John didn’t call someone.
I am surprised that John didn’t call someone.
I regret that John didn’t call someone.
If we don’t call someone, we are doomed.
Every boy who didn’t call someone . . .
Only John didn’t call someone.
Few boys didn’t call someone.
√ not > not > some
√ no one > not > some
√ surprise > not > some
√ regret > not > some
√ if (not > some)
√ every (not > some)
√ only > not > some
√ few > not > some
Why would a PPI under negation become legitimate in NPI contexts? Szabolsci suggests that this
is so because PPI plus negation is an NPI itself. PPIs are claimed to “have two NPI-features. One
is a strong-NPI feature like that of yet and squat: it requires a clausemate antiadditive licensor,
without intervention. The other is a weak-NPI feature like that of ever: it requires a Strawsondecreasing licensor (not necessarily clausemate but without intervention)…. I propose that these
6
van der Wouden (1994) also identifies a class of bipolar items: these are claimed to require a decreasing licensor
(an NPI-property) but cannot occur under a local antimorphic item (he calls this a PPI-property). Van der Wouden
argues that NPI-hood and PPI-hood are two primitive properties and may therefore coexist in one item.
35
two features are normally ‘dormant’. A context that can license the strong-NPI feature ‘activates’
and, in the same breath, licenses that feature. What we have seen indicates, however, that the
other, weak-NPI feature also gets activated at the same time – activated, but not licensed.
Therefore, the emergent constellation is illegitimate, unless a licensor for the weak-NPI feature is
provided. In other words, PPIs do not detest antiadditives; they have a latent craving for
antiadditives. That they appear to detest them is due to the fact that the satisfaction of this
craving activates another, which needs to be satisfied independently.” (Szabolcsi 2004: 429).
In such an account, a negative condition (anti-licensing) is reduced to a positive one
(licensing), and an underlying NPI source is posited in both NPIs and PPIs. The exact nature of
the commonallity needs to be refined, but the appeal of the reasoning here cannot go unnoticed.
However, there are reasons to be skeptical about going that route. One obvious obstacle is that
this account envisions NPI and PPI licensing in terms of syntactic features purely—negations, in
particular— and gives us little insight into the lexical semantics of some itself.
Szabolcsi is correct to point out that some is not scalar, or strictly referential: clearly, in the
cases above it takes narrow scope. Szabolcsi is also right to point out that we cannot extend the
referentiality approach as to all PPIs— e.g. would rather, already, and speaker oriented adverbs are
obviously not amenable to a referential analysis. (Given the variable sources of NPI sensitivity that
we must allow for, there is no reason to a priori expect the PPI domain to lack a comparable
variation). However, Szabolcsi treates some as ¬¬∃ , with the two negations canceling each other
out, and it is difficult to see this as more than mere stipulation. What is the evidence for the two
negations? And why do we never see overt realizations of them in some crosslinguistically? This
is typologically quite surprising, because negation in languages is never “forgotten” to be
marked, if there. Ultimately, why items like ¬¬∃ exist? Why would a language bother to
implement two negations on an expression just to cancel them out?
In assessing Szabolci’s data, it is important to note two things. First, the narrow scope some
in (33-41) is non-emphatic; reproducing the examples with emphatic SOME is either odd, or
allows a wide scope reading:
(130)
a
b
c
d
e
f
# I don’t think that John didn’t call SOMEONE.
# No one thinks that John didn’t call SOMEONE.
# I am surprised that John didn’t call SOMEONE.
# Every boy who didn’t call SOMEONE . . .
# Only John didn’t call SOMEONE
# Few boys didn’t call SOMEONE.
The judgments here must be checked carefully, and a large scale inquiry is needed to establish
the conditions on the availability of the two intonational patters for some. What is crucial,
however, is that the narrow scope correlates with non-emphatic intonation, and that nonemphatic
intonation is not the intonation observed with clausemate negation. Based on this contrast, it
becomes plausible to argue that the narrow scope some is a different species from the PPI
emphatic SOME under negation, and by different species, I mean lexically distinct. Intonation
functions as a morphological feature, something not at all peculiar, but familiar from emphatic
non-emphatic pairs generally observed with negation (Giannakidou 1998, 2000, 2006, and works
cited there; Yoshimura 2007). In the observed cases, the emphatic member of the pair is the one
that outscopes negation. Emphatic some, then, is simply part of this general picture.
A second observation, missed by Szabolcsi, is that lower negation is in fact not necessary:
narrow scope non-emphatic some is “licensed” without it:
36
(131)
a
b
c
d
e
f
g
h
I don’t think that John called someone.
No one thinks that John called someone.
I am surprised that John called someone.
I regret that I called someone.
If we call someone, we’ll get help.
Every boy who called someone . . .
Only John called someone.
Few boys called someone.
Non-clausemate negation some is part of this pattern too: *John didn’t see that he talked to
SOMEONE is very odd. Nonemphatic some thus appears in polarity contexts generally, and does
not dependent on the presence of negation; in fact, it appears freely in any context, regardless of
veridicality or monotonicity, as indicated below.
(132)
a
b
c
Billed talked to someone (but I don’t know who).
Bill want to talk to someone (but I don’t know who).
Every student visited some museum (but I don’t know which ones).
Hence Szabolcsi’s narrow scope some is nothing more than the “regular” indefinite some, which
can be specific or non-specific depending on the context. There seems be nothing more to
explain in these cases. I think this is indeed the right way of looking at this problem: there are
two varieties of some indefinites, an emphatic and a non-emphatic some, and only emphatic
SOME is polarity sensitive. Non-emphatic some is a regular indefinite with no restrictions in its
distribution. One would expect the intonation strategy to be systematically available across
languages—an issue to investigate further in the future.
So, why does emphatic SOME need to scope above negation? The most profitable avenue
will be to think of SOME as an indefinite that can only be interpreted referentially. In this sense,
SOME is the opposite of dependent indefinites that we identified earlier, and which can never be
interpreted referentially. SOME can be seen as an indefinite that is always specific—say, a
variant of a certain, or a particular. Notice their parallel wide scope with negation:
(133)
a
b
c
Sue didn’t talk to a certain Norwegian—his name is Otto.
Sue didn’t talk to a particular Norwegian—his name is Otto.
Sue didn’t talk to SOME Norwegian—his name is Otto.
A certain, a particular, and SOME all want to escape negation. We can argue that intonation
provides an abstract morphological feature, e.g. equivalent to certain, or specific, that turns some
specific and forces it to move to a scope position above negation. For reasons of space, I will not
develop this suggestion further—and the first step is to examine whether SOME in contexts
outside negation behaves on a par with a certain and a particular. This would be the hope, and
the typical case of scope interaction with a universal quantifier below allows us to be optimistic:
(134)
Every student visited SOME museum. (a specific museum)
The advantage of viewing the wide scope property of PPI-SOME as an instance of
37
specificity is obvious: SOME is no longer mysterious, but a member of the class of wide scope
indefinites that is fairly well identified (though perhaps not as well understood; see Farkas 2002
for an overview). Szabolcsi’s account, by contrast, fails to capture this link. Instead, PPI-some
comes out as a very strange expression, a indefinite with two negations attached to it, for which
there is no obvious motivation. In that account, the uniform pattern between a certain, a
particular, and SOME remains accidental.
8.2 Speaker oriented adverbs as PPIs
Speaker oriented (unfortunately) and modal (possibly) adverbs have been analyzed recently as
polarity items by Nilsen (2003) and Ernst (2008). The main observation here too is that these
adverbs are incompatible with the scope of local negation:
(135)
a
b
John unfortunately disappeared.
#John didn’t unfortunately disappear.
The positive sentence says that John disappeared and that this is unfortunate for the speaker. The
negation of this sentence ought to express truth reversal: John did not disappear and this is not
unfortunate. Rather than saying this, however, the negative sentence comes out odd.
Unfortunately is a PPI, Ernst argues, because it expresses strong subjectivity. Strong
subjectivity means that the adverb is factive-like: upon uttering (135a), the speaker is committed
to the truth of John disappearing, and further asserts that this is unfortunate. This means that
unfortunately is veridical:
(136) [[ John has unfortunately disappeared]]c= 1 iff
∀w [w ∈ ME (speaker) → w ∈ λw'. John disappeared in w']
Every world in the speaker’s epistemic model is a world where John disappeared. From this,
incompatibility with negation follows: negation would require that the proposition be false in all
the worlds, and this leads to a contradiction. This type of reasoning predicts oddity and not
ungrammaticality, and this is precisely the status that (135b) with illicit unfortunately.
Other nonveridical sentences— questions, conditionals are also odd with unfortunately:
(137)
a
B
#Has he unfortunately disappeared?
#If he has unfortunately disappeared…
Nonveridical contexts allow some worlds in ME (speaker) to not be p-worlds, and this again
would lead to a contradiction, given the truth condition (136) of unfortunately. Factive adverbs
like unfortunately, then, are veridical and will only be usable in veridical contexts.
Ernst further shows that there is variation within the adverb PPI class—unfortunately is
excluded from all nonveridical contexts, but modals (possibly), and what he calls weak
evaluatives (mysteriously) can appear in questions and the antecedent of conditionals given
certain conditions. Below, I give Ernst’s chartwhich summarizes the variation, and some
examples from his paper to illustrate:
38
(138) Variation in PPI adverbs
(Ernst 2008: Table 1)
Adverb
Regular
Questions/ Negative
Negative
Low-tone
Other
type
negation
condiquestions
counterdenial
metalinguistic
tionals
factuals
MN
negation (MN)
---------------------------------------------------------------------------------------------------------------a. Strong
*
*
*
*
*
OK
evaluatives
(unfortunately, luckily)
b. Weak
*
*/OK
OK
OK
OK
OK
evaluatives
(mysteriously, conveniently)
c. Modals
*
*/OK
?/OK
*/OK
OK
OK
(probably, possibly)
d. Evidentials
OK
OK
OK
OK
OK
OK
(clearly, obviously)
---------------------------------------------------------------------------------------------------------------(139) a. Are they probably going to be invited to the meeting?
b. Where have they probably put the loot?
c #Are they unbelievably going to be invited to the meeting?
(140) a. If, as you say, they're probably in line for an award, maybe we should get tickets for
the ceremony as soon as we can.
b. If they have conveniently decided to withdraw, the competition will go better for us.
c. # If they have luckily decided to withdraw, the competition will go better for us.
Notice that strong evaluatives remain consistently odd in questions, conditionals. This contrast
and variation, unnoticed in Nilsen (2003), suggests that not all speaker oriented adverbs express
full speaker commitment; mysteriously, probably express partial commitment, hence the truth of
p does not hold in all worlds in ME (speaker), but in a subset of this model. This predicts
incompatibility with negation, but greater flexibility with respect to nonveridical operators.
Evidentials, on the other hand, are objective: they rely on evidence for the truth of p outside the
speaker’s beliefs. If speaker commitment is the source of PPI-status, we thus capture nicely the
fact that evidential adverbs are not PPIs (and are fine with negation):
(141)
John didn’t clearly express his desires.
Space prevents me from elaborating more on these very interesting ideas (see Ernst’s paper for
more details). The two important things to emphasize is the variation in the PPI domain—a
welcome result, given the variation we attest in the NPI domain—and that speaker commitment
is the key to understanding the incompatibility of speaker oriented adverbial PPIs with negation.
One final point worth highlighting is that scalarity is not relevant for speaker oriented
adverbs either. I am not going to repeat Ernst’s arguments against Nilsen’s scalar analysis of
possibly (which stipulates the reverse of widening: domain shrinking). Ernst’s main objections
are that there is no evidence for shrinking, that this property is stipulated as a lexical feature for
theory internal reasons only, and that it misses the central connection to speaker commitment.
39
The nonveridicality approach captures precisely this link, and uses it to account for the PPI
nature of speaker oriented adverbs, and the variation attested in this class.
9
Conclusion
The main conclusions to be drawn from our discussion of NPIs and PPIs in this paper are the
following. First, polarity patterns crosslinguistically reveal two main kinds of sensitivity: (a)
more local sensitivity to anti-veridicality (negation, antiadditivity, and antimorphicity included),
and (b) a broader sensitivity to nonveridicality. The former characterizes a class of NPIs that
have a very strict distribution, and is often realized also as a syntactic dependency (agreement or
QR). Anti-veridicality also explains the incompatibility of referential PPIs (emphatic SOME)
with negation. Importantly, we have seen that mere downward entailment is only a very weak
NPI licenser, and often, not a licenser at all (Greek NPIs of various types, Dutch ook maar, and
similar items; recall also that at most, and even few, often give variable judgments in English
too). For PPIs, mere DE plays no role (Szabolcsi 2004, Ernst 2008).
We also found that scalar approaches to NPIs cannot provide a conceptually or analytically
secure foundation for explaining why NPIs appear in nonveridical contexts. Such approaches
predict the wrong status of illicit NPIs (merely uninformative or contradictory, when NPIs are in
fact ungrammatical), and they are unable to capture the variation in the distribution of scalar PIs:
e.g. why some of them are good with negation (NPIs), but some others are not (FCIs). Finally,
the scalar approach does not seem a plausible theory for NPIs that are not scalar, or for PPIs,
which as a class seems to be non-scalar.
I suggested that it is empirically and analytically more attractive to think of polarity
phenomena as diverse in specific ways. Polarity items are created because of distinctive lexical
properties they have, and we can identify three sources of polarity sensitivity: (a) EVEN-based
scalarity (with four possible meanings of EVEN; Giannakidou 2007), (b) a referential deficiency
that creates non-deictic variables that cannot be interpreted as free variables, and (c) degrees of
speaker commitment (for speaker oriented PPI adverbs; Ernst 2008). Such an approach is more
consistent with the empirical and interpretational diversity of PIs, and offers a flexible enough
framework to address compositionally the individual sensitivities of the various NPI and PPI
classes, and predict their correct grammatical status.
Acknowledgment
I would like to thank the editors of the volume for giving the opportunity to write this article, and
Paul Portner in particular for making sure that I finish it. Since I started working on this paper, I
presented parts of the material at the Zentrum fur Allgemeine Sprachwissenschaft (ZAS) Berlin,
UMass Amherst Linguistics Colloquium, and the Semantics Research Seminar at Keio
University in Tokyo, Japan. I would mike to thank the audiences for their very useful feedback. I
also benefited enormously from discussions with Jay Atlas, Jack Hoeksema, Larry Horn,
Angelika Kratzer, Manfred Krifka, Jason Merchant, Barbara Partee, Chris Potts, Chris Tancredi,
Yoad Winter, and Frans Zwarts. Many thanks to all for your extremely helpful comments! The
work left to be done will hopefully be not entirely my responsibility.
References
40
Atlas, Jay David. 1993. The importance of being Only: Testing the Neo-gricean versus Neoentailment paradigms. Journal of Semantics 10.301-318.
Atlas, Jay David. 1996. Only Noun Phrases, Pseudo-negative quantifiers, negative polarity items,
and monotonicity. Journal of Semantics 13. 265-328.
Baker, Carl Lee. 1970. ‘Double Negatives’, Linguistic Inquiry 1: 169–186.
Bernardi, Raffaella. 2002. Reasoning with Polarity in Categorial Type Logic. PhD Diss.
University of Utrecht.
Beaver, David, and Brady Clark. 2003. Always and only: why not all focus sensitive operators
are alike. Natural Language Semantics 11.323-362.
Borschev, V., E. Paduscheva, A. Partee, Y. Testelets, and I. Yanovich. 2007. Russian genitives,
non-referentiality, and the property-type hypothesis. To appear in Proceedings of FASL.
Cheng, Lisa L.-S. and Anastasia Giannakidou to appear. The Non-Uniformity of whindeterminates with free choice in Chinese. To appear in Georges Tsoulas and Kook-hee
Gil (eds.), Strategies of Quantification, Oxford University Press.
Chierchia, Gennaro. 2006. Broaden your views. Implicatures of domain widening and the
“Logicality” of language. Linguistic Inquiry 37: 535-590.
Dayal, Veneeta (1998), Any as inherently modal. Linguistics and Philosophy 21:433-76.
Dowty, David. 1994. ‘The Role of Negative Polarity and Concord Marking in Natural Language
Reasoning’, in Harvey and Santelmann (eds.), SALT IV, Cornell University, Ithaca, New
York, pp. 114–145.
Drenhaus, Heiner, beim Graben, P., Frisch, S., and D. Saddy. 2006. Diagnosis and repair of
negative polarity constructions in the light of symbolic resonance analysis. Brain and
Language 96:255-268.
Drenhaus, Heiner, Joanna Blaszczak, and Juliane Schutte. 2006. An ERP study on the strength of
licensers in negative polarity constructions. Under submission.
Dikken, Marcel den, and Anastasia Giannakidou (2002), From Hell to Polarity: ‘Aggressively
non-D-linked Wh-phrases as polarity items’. Linguistic Inquiry 33:31-61.
Ernst, Thomas. 2008. Speaker oriented adverbs. (Sumbitted) manuscript, Dartmouth
College/UMass Amherst.
Farkas, Donka F. 2002. Varieties of specificity. Journal of Semantics.
von Fintel, Kai. 1999. NPI-Licensing, Strawson-Entailment, and Context-Dependency. Journal
of Semantics 16.97-148.
von Fintel, Kai (2000), Whatever. Proceedings of SALT 10: 27-40.
Giannakidou, Anastasia. 1994. “The semantic licensing of NPIs and the Modern Greek
subjunctive”. In Language and Cognition 4, Yearbook of the Research Group for
Theoretical and Experimental Linguistics: 55-68. University of Groningen.
Giannakidou, Anastasia. 1995. Subjunctive, habituality and negative polarity. Semantics and
Linguistic Theory (SALT) V: 132-150. Cornell University, Ithaca, NY.
Giannakidou, Anastasia 1997. The Landscape of Polarity Items. PhD thesis, University of
Groningen.
Giannakidou, Anastasia (1998), Polarity Sensitivity as (Non)veridical Dependency. John
Benjamins. Amsterdam.
Giannakidou, Anastasia (1999). Affective dependencies. Linguistics and Philosophy 22: 367421.
Giannakidou, Anastasia (2001), The meaning of free choice. Linguistics and Philosophy 24:
659-735.
41
Giannakidou, Anastasia. 2002. Licensing and sensitivity in polarity items: from downward
entailment to nonveridicality. In Maria Andronis, Anne Pycha and Keiko Yoshimura
(eds), CLS 38: Papers from the 38th Annual Meeting of the Chicago Linguistic Society
Giannakidou, Anastasia, 2003. EVEN in questions revisited. In the Proceedings of the 14th
Amsterdam Colloquium, ed. by Paul Dekker and Martin Stokhof. Institute for Logic,
Language, and Computation, University of Amsterdam. 56-62.
Giannakidou, Anastasia (2006), Only, emotive factive verbs, and the dual nature of polarity
dependency. Language 82: 575-603.
Giannakidou, Anastasia. 2007. The landscape of EVEN. Natural Language and Linguistic
Theory.
Giannakidou, Anastasia, to appear. The dependency of the subjunctive revisited: temporal
semantics and polarity. Lingua, special issue on Mood.
Giannakidou, Anastasia and Lisa Cheng. 2006. (In)definiteness, polarity, and the role of whmorphology in free choice. Journal of Semantics 23:135-183.
Giannakidou, Anastasia and Josep Quer. Long distance licensing of negative indefinite. In
Negation and Polarity, Syntax and Semantics, ed. By D. Forget et al. John Benjamins. 95113.
Guerzoni, Elena and Yael Sharvit. 2007. A question of strength: on NPIs in interrogative clauses.
Linguistics and Philosophy 30: 361-391.
Haspelmath, Martin (1997), Indefinite Pronouns. Oxford University Press. Oxford.
Herburger, Elena (2003). A note on Spanish ni siquiera, even, and the analysis of NPIs. Probus
15, 237-256.
Herburger, E. and S. Mauck. 2006. NPIs are low scalar items +F. Ms. Georgetown University.
Hoeksema, Jacob. 1986. Monotonie en superlatieven. Proeven van taalwetenschap, ed. by C.
Hoppenbrouwers, Joop Houtman, Inneke Schuurman, and Frans Zwarts. Groningen:
Nederlands Instituut RUG. 38-49.
Hooksema, 1995. Ooit.
Heim, Irene (1982), The Semantics of Definite and Indefinite Noun Phrases. Unpublished Ph.D.
dissertation, University of Massachussets. Amherst, MA.
Heim 1984. A note on negative polarity and downward entailingness. In NELS 14: 98-107.
Horn, L. 1972. On the semantic properties of Logical operators in English. PhD thesis. UCLA.
Horn, Laurence. 1989. A Natural History of Negation, Chicago University Press.
Horn, Laurence R. (1997), All John’s Children are as Bald as the King of France: existential
import and the geometry of opposition. Chicago Linguistics Society 33:155-179.
Horn, Laurence R. (2000a), Any and (-)ever: Free choice and free relatives. Proceedings of
Israeli Association for Theoretical Linguistics 15:71-111.
Horn, Laurence R. (2000b), Pick a theory: not just any theory. In L. Horn & Y. Kato (eds.),
Negation and Polarity: Syntactic and Semantic Perspectives. Oxford University Press.
Oxford. 147-192.
Horn, Laurence R. 2002. Assertoric inertia and NPI-licensing. Chicago Linguistics Society 38:
Parassesion on Negation and Polarity, 55-82.
Horn, Laurence R. (2006b), Airport '68 Revisited: Toward a unified indefinite any. In Gregory
Carlson and F. Jeffrey Pelletier (eds.), The Partee Effect. CSLI. Stanford. 179-205.
Horn, L.R. 2006b. The boarder wars: a neo-Gricean perspective. In Where Semantics Meets
Pragmatics: Current Research in the Semantics/Pragmatics iInterface, Vol. 16, ed.by K.
von Heusinger and K. Turner, 21-48. Elsevier.
42
Israel, Michael (1996). Polarity sensitivity as lexical semantics. Linguistics and Philosophy 19,
619-666.
Israel 2004. The pragmatics of polarity. In The Handbook of Pragmatics, Horn, L. and G. Ward
(eds), Blackwell. pp. 701-723.
Jackson, Eric. 1994. Negative Polarity, Definites under Quantification and General
Statements. PhD Dissertation, Stanford University.
Jacobson, Pauline. 1995. On the quantificational force of the English free relative. In Bach et al.,
Quantification in Natural Language. Kluwer. 45-487.
Jespersen, Otto. 1909–1949. A Modern English Grammar on Historical Principles, George,
Allen, and Unwin Ltd., London.
Jayez, J. and L. Tovena. 2005. Free choiceness and non-individuation. Linguistics and
Philosophy 28: 1-71.
Kadmon, Nirit and Fred Landman (1993), Any. Linguistics and Philosophy 16:353-422.
Kay, Paul (1990). Even. Linguistics and Philosophy 13, 59-111.
Kratzer, Angelika, and Shimoyama, Junko. 2002. Indeterminate Phrases: the View from
Japanese. In The Proceedings of the Third Tokyo Conference on Psycholinguistics, 1-25.
Tokyo: Hituzi Syobo.
Krifka, Manfred (1995), The semantics and pragmatics of polarity items in assertion. Linguistic
Analysis 15:209-257.
Krifka, Manfred, F. J. Pelletier, Greg Carlson, A. ter Meulen, G. Chierchia & G. Link (1995),
Genericity: an introduction. In G. Carlson & F. J. Pelletier (eds.), The Generic Book.
University of Chicago Press. Chicago. 1-124.
Ladusaw, William. 1980. Polarity Sensitivity as Inherent Scope Relations. Garland Publishing.
Ladusaw, William. 1983. “Logical Form and Conditions on Grammaticality.” Linguistics and
Philosophy 6: 373-92.
Ladusaw, William. 1996. “Negation and Polarity Items.” The Handbook of Contemporary
Semantic Theory. Shalom Lappin, ed., 321-341. Oxford and Malden: Blackwell.
Lahiri, Utpal. 1998. Focus and negative polarity in Hindi. Natural Language Semantics 6.57-23.
Laka, Itziar. 1990. Negation in Syntax: on the Nature of functional categories and projections.
MIT dissertation.
Lee, Chungmin (1997), Negative polarity and free choice: where do they come from? In P. Dekker et al. (eds.),
Proceedings of the 11th Amsterdam Colloquium. ILLC, University of Amsterdam. 217-222.
Lee, Jung-Huyck. 2007. Korean EVEN: -to versus -(i)lato. Manuscript, University of Chicago.
Lee, Young-Suk and Laurence R. Horn (1994), Any as an indefinite plus even. Unpublished MS. Yale
University.
LeGrand, Jean Ehrenkranz (1975), Or and Any: The semantics and syntax of Two logical
Operators. Unpublished Ph.D. dissertation. University of Chicago.
Levinson, Dmitry. 2006. Polarity Sensitivity in Inflectional Morphology. Paper presented at
Workshop on Polarity, Swarthmore College.
Lin, Jo-Wang. 1996. Polarity Licensing and Wh-phrase Quantification in Chinese, PhD
dissertation, University of Massachusetts, Amherst.
Linebarger, Marcia. 1980. The Grammar of Negative Polarity. PhD thesis. MIT.
Linebarger, Marcia 1987. Linebarger, Marcia. 1987. ‘Negative Polarity and Grammatical
Representation’, Linguistics and Philosophy 10, 325–387.
Linebarger, MArcia. 1991. Negative polarity as lingusitic evidence. CLS 27, PArasession on
Negation. 165-188.
43
Matthewson, Lisa. 1998. Determiner Systems and Quantificational Strategies. Evidence from
Salish. Worlds’s theses 1. Holland Academic Graphics.
McNally, Louise. 1998. Existential sentences without existential quantification. Linguistics and
Philosophy 21:353-392.
Montague, Richard. 1969. On the nature of certain philosophical entities. The Monist 53: 159-94.
Nakanishi, Kimiko. 2007. Presentation at the LSA 2007 Meeting, Workshop on Polarity.
Nathan, Lance. 1999. Either: Negative Polarity Meets Focus Sensitivity. Honors thesis. Brown
University.
Nilsen, Oystein. 2003. Eliminating Positions. PhD thesis. University of Utrecht.
Partee, B. 1986. Noun phrase interpretation and type shifting prinviples. In Studies in Discourse
Representation Theory and the Theory of Generalized Quantifiers. Dordrecht. 115-143.
Partee, B. and V. Borschev, 2004. The semantics of Russian genitive of negation: the role and
nature of perspectival structure. In SALT 14: 212-234.
Pereltsvaig, Asya. 2000. Monotonicity-based versus veridicality based approaches to negative
Polarity: evidence from Russian. Formal Approaches to Slavic Linguistics, ed. by Tracy
Halloway King, and Anna Sekerina, 328-346. Ann Arbor: Michigan Slavic Publishers
Potts, Christopher. 2005. The Logic of Conventional Implicatures. Oxford University Press.
Progovac, Ljiljana. 1994. Positive and Negative polarity: a binding approach. Cambridge
University Press.
Progovac, Ljiljana. 2000. ‘Negative and Positive Feature Checking and the Distribution of
Polarity Items’, In Sue Brown and A. Przepiorkowski (eds.), Negation in Slavic,
Indiana University, Bloomington.
Przepiórkowski, Adam and Anna Kupc . 1998. ‘Eventuality Negation and Negative Concord in
Polish and Italian’, in Robert D. Borsley and Adam Przepiórkowski (eds.), Slavic in
HPSG, CSLI, Stanford University, pp. 211–246.
Postal, Paul. 2000. The Ohio Lectures on squat. Manuscript. New York University, New York.
Richter, Frank and Manfred Sailer. 1998. ‘LF Constraints on Expressions of Ty2: An HPSG
Analysis of Negative Concord in Polish’, in Robert Borsley and Adam Przepiórkowski
(eds.), Slavic in HPSG, CSLI Stanford University, pp. 247–282.
Van Rooy, Robert. 2003. Negative polarity items in questions: strength as relevance. Journal of
Semantics 20. 239-273.
Rullmann, Hotze. (1996). Two types of negative polarity items. In K. Kusumoto et al. (eds).
Proceedings of NELS 26. GLSA, 335-350.
Rullmann, Hotze (1997). Even, polarity, and scope. In M. Gibson, G. Wiebe, and G. Libben
(eds). Papers in Experimental and Theoretical Linguistics. University of Alberta,
Edmonton.
Rullmann, Hotze. 2003. Additive particles and polarity. Journal of Semantics 20: 329-401.
Quer, Josep (1998), Mood at the Interface. Ph.D. dissertation, University of Utrecht.
Quer, Josep (1999), The quantificational force of free choice items. Unpublished MS. University
of Amsterdam.
Rooth, Mats. 1985. Association with Focus. PhD thesis, UMass.
Russell, Benjamin. 2007. Against Grammatical Computation of Scalar Implicatures. Journal of
Semantics.
Seuren, Pieter. 1984. The comparative revisites. Journal of Semantics 3:109-141.
Saddy, D., Drenhaus H., and S. Frisch, 2004. Processing polarity items: contrastive licensing
costs. Brain and Language 90: 405-502.
44
Shimoyama. Junko. 2003. Wide scope universals in Japanese. Ms. McGill University.
Stalnaker, Robert. 1978. Assertion. In Syntax and semantics 9: Pragmatics, ed. by Peter Cole,
315 – 332. New York: Academic Press.
Suranyi, Balazs. 2006. Quantification and focus in negative concord. Lingua 116: 272-313.
Szabolcsi, Anna. 2004. Positive polarity-Negative polarity. Natural Language and Linguistic
Theory 22.409-452.
Tancredi, Chris. 2007a. A multi-model modal theory of I-semantics, Part I: Modals. Univesity of
Tokyo.
Tancredi, Christopher, 2007b. A multi-model modal theory of I-semantics, Part II: Identity and
Attitudes.
Tovena, Lucia. 1998. The Fine Structure of Polarity Items. Garland, NY.
van der Wouden, Ton. 1994. Negative Contexts. PhD thesis, University of Groningen.
Uribe, Etxebarria, Miriam. 1994. Interface Licensing Conditions on Negative Polarity Licensing:
A theory of Polarity and Tense Interactions. PhD thesis. University of Connecticut.
Vlachou, Evangelia. 2007. Free Choice Items In and Out of Context. PhD thesis. University of
Utrecht.
Yoshimura, Keiko. 2007. Focus and Polarity in Japanese. PhD thesis, University of Chicago.
Zwarts, Frans. 1981. ‘Negatief polaire uitdrukkingen 1’, GLOT 4, 35–132.
Zwarts, Frans. 1986. Categoriale Grammatica en Algebraische Semantiek. Een studie naar
negatie en polariteit in het Nederlands. PhD thesis, University of Groningen.
Zwarts, Frans. 1993. Three types of polarity. Ms. Also appeared as Zwarts 1998 in Plural
Quantification, ed. F. Hamn and E. Hinrichs, Kluwer, Dordrecht.
Zwarts, Frans. 1995. Nonveridical contexts. Linguistic Analysis 25. 286-312.
45